cli

package module
v0.1.92 Latest Latest
Warning

This package is not in the latest version of its module.

 Go to latest Published: Aug 10, 2026 License: Apache-2.0 Imports: 1 Imported by: 0

README ΒΆ

civitai CLI

Browse and download Civitai models, images, and articles β€” and author, validate, and submit App Blocks. Two paths in one static binary: a read/download client for the public API (reads are anonymous; downloads need a token), and the toolchain for shipping Apps (every civitai app command needs one).

⚠️ Apps is in a limited, invite-only beta (pre-GA). You can install this CLI, login, scaffold, validate, and run an app locally right now β€” but civitai app submit and dev:live require an invite: submission and dev:live are limited to invited beta testers while the feature is in a limited (pre-GA) beta, until Apps opens to the public.

Anyone can request an invite β€” open a request below and we'll review it:

Request access

The command-line interface for Civitai β€” a single static binary that does two things: it's a thin read/download client for Civitai's public API (browse and fetch models, images, and articles β€” no account needed to read those), and it's the toolchain to author, validate, and ship Apps. Everything under civitai app needs a credential, including the App-store browse commands app list / app view β€” see Browse the App store.

An App is a small, sandboxed web app that runs inside Civitai surfaces (it's served in an iframe; the platform owns the build and the runtime). The CLI replaces the error-prone "hand-format a ZIP" flow: it scaffolds a correct project, validates the manifest against the platform contract, and packages/submits it for review.

New here? The Build your first App guide is the full end-to-end walkthrough.

Contents

Get started

Author an App

Use the API

Reference

Install

Pick whichever fits β€” npm is the most convenient if you already have Node (App authors usually do); Homebrew is quickest on macOS/Linux; the prebuilt binary needs no toolchain; go install builds from source.

npm (Node)

A thin wrapper that downloads the matching prebuilt binary for your OS/arch on install and verifies its sha256 against the release checksums.txt:

npm install -g @civitai/cli
# or run it without installing:
npx @civitai/cli --help
Homebrew (macOS / Linux)
brew install civitai/tap/civitai
Nix flake

This repo is a Nix flake, so you can run or install civitai without a Go toolchain (works on x86_64/aarch64 Linux and macOS):

# Run without installing:
nix run github:civitai/cli -- models search "sdxl"

# Install into your Nix profile:
nix profile install github:civitai/cli

Pin it as an input in your own flake:

{
  inputs.civitai-cli.url = "github:civitai/cli";

  outputs = { self, nixpkgs, civitai-cli }: {
    # e.g. add to a devShell / home-manager / systemPackages:
    #   civitai-cli.packages.${system}.default
  };
}
Prebuilt binary

Download a prebuilt binary for your OS/arch from the GitHub Releases page (linux, macOS, windows Γ— amd64/arm64), verify it against checksums.txt, then put it on your PATH:

tar xzf civitai_*_linux_amd64.tar.gz
sudo mv civitai /usr/local/bin/
civitai version
Go install (from source, Go 1.25+)
go install github.com/civitai/cli/cmd/civitai@latest
# installs the `civitai` binary into $(go env GOPATH)/bin

Quickstart: browse & download

Reads of the public catalog β€” models, model versions, images, tags, creators, users, articles, collections β€” are anonymous: no login needed for the commands in this section. Every one of them takes --json to emit the raw API response for scripting.

What is not anonymous. civitai download needs a token, and so does every civitai app … command β€” including the App-store browse commands civitai app list and civitai app view, which exit 3 with no token configured when you have not logged in. The store endpoint keys the visible catalog off your identity, so there is no anonymous view of it. See Browse the App store.

# Search models β€” filter by base model, type, and sort:
civitai models search --base-model Illustrious --type Checkpoint --sort "Most Downloaded"

# --base-model works on any type, including embeddings (TextualInversion):
civitai models search --type TextualInversion --base-model "SDXL 1.0"

# Inspect a specific model or a specific model version:
civitai models get 618692
civitai model-versions get 691639

# Download a version's file(s) β€” SHA256-verified, streamed atomically.
# `--layout` routes each file into the right app subfolder (also `a1111`);
# `--dry-run` prints the plan without transferring. Downloads require `civitai login`.
civitai download 691639 --layout comfyui --root ~/ComfyUI
civitai download 691639 --dry-run

# Find and read articles (guides) right in the terminal:
civitai articles search --query "comfyui workflow"
civitai articles get 32680 --content

See Browse the public API and Download model files below for the full command and flag reference (images, tags, creators, collections, pagination, folder routing, base-model compatibility checks, and more).

Quickstart: build an App Block

# 1. Authenticate once (browser device login; or `civitai login --token <t>`).
civitai login

# 2. Scaffold a ready-to-build App (batteries-included page-money default).
civitai app create my-app
cd my-app

# 3. Install deps and run it locally against the mock host (no real Buzz/compute).
#    `npm run dev` alone renders blank β€” the harness supplies the host.
npm install
npm run dev:harness

# 4. Edit your app; build it, then check the manifest before submitting.
#    (the `static` template has no build step β€” skip `npm run build`.)
npm run build
civitai app validate

# 5. Package + submit for review (uploads with your stored token by default).
#    Interactively this asks you to confirm. In CI β€” or any non-TTY shell β€”
#    a token-carrying submit REFUSES without --yes rather than firing a real
#    moderator-review request nobody approved. Scripts must pass it:
#      civitai app submit --yes
civitai app submit

# 6. Check where your submission is in review / deploy.
civitai app status

# 7. Attach the store-listing media. An icon AND a cover are REQUIRED before the
#    listing can publish β€” do it now, while the app is in review; it carries
#    forward on approval. `listing status` shows what's still missing.
#    The scaffold creates `assets/` and a README of the requirements, but NO
#    images β€” save your own icon.png and cover.png in there first:
#      icon   png/jpeg/webp, <= 2 MiB, square-ish  β€” start from 512 x 512
#      cover  png/jpeg/webp, <= 4 MiB, landscape   β€” start from 1600 x 900
#    Full bounds (and who checks what) in "Listing media requirements" below.
civitai app listing set-icon ./assets/icon.png
civitai app listing set-cover ./assets/cover.png
civitai app listing status

Step 7 needs artwork you supply. Every template scaffolds an assets/ directory with a README of the requirements, and deliberately no placeholder images β€” a placeholder passes every check and uploads cleanly, which is how a stub icon reaches a public listing. Sizes, formats and aspect ratios are in Listing media requirements.

Want to drive the real backend (real Buzz/compute) before submitting? Mint a dev token with civitai app dev-token and run npm run dev:live β€” see Local dev loop.

Submit β‰  live. civitai app submit enters your app into moderator review β€” it is not published immediately. The lifecycle is submit β†’ review β†’ approve β†’ build + deploy β†’ https://<blockId>.civit.ai/: that URL 404s until a moderator approves your submission and the platform builds + deploys it (a few minutes after approval). Until then, track status on /apps/my-submissions (a fresh submission sits at pending). See Submit & auth for the full flow. (And note Apps is in an invite-only beta β€” see the warning above.)

Enable shell completion (optional):

source <(civitai completion bash)   # bash; see `civitai completion --help` for zsh/fish/powershell

SDK packages

This CLI scaffolds, validates, and submits β€” but the code your app actually imports lives in two published npm packages (the page-money template wires them for you; static and page-vite are deliberately dependency-free):

Package What it is
@civitai/blocks-react The React hooks + iframe transport app authors call β€” useBlockContext, useBuzzWorkflow, useBlockResize, the /ui component pack, and the /testing dev hosts. Start here for the hook reference.
@civitai/app-sdk The framework-agnostic contract under the hooks β€” manifest types, scope strings, the postMessage protocol, and the defineBlock validator (@civitai/app-sdk/blocks). 
# Already installed by the scaffold; this is the explicit install line:
pnpm add @civitai/blocks-react @civitai/app-sdk react

The full hook-by-hook reference (with snippets) lives in each package's npm README. For the end-to-end walkthrough, see Build your first App.

Command reference

Command What it does
civitai login [--scopes <set>] [--token [<t>]] [--no-browser] Browser OAuth device login by default (stores auto-refreshing tokens). The default scope set grants identity + Apps submit + dev-tunnel and not Buzz-spend; --scopes generate additively grants generation + Buzz spend (needed by civitai generate and money-path dev:live). --token <t> stores a personal API key instead (not combinable with --scopes). --token with no value prints where to create a personal key (civitai.com/user/account) and how to re-run β€” handy when you know you want a personal key but haven't minted one yet. Config at ~/.config/civitai/config.yaml, 0600. Also reads CIVITAI_TOKEN.
civitai whoami [--scopes] [--json] Verify the stored token; print the authenticated user and a Capabilities section β€” credential type (OAuth login vs personal API key), Read Buzz balance, and Spend Buzz β€” decoded from the token's scope, so a money-path dead end (a default OAuth login can't spend) is visible before dev:live β€” and when it can't, the output names the fix for that credential (login --scopes generate for an OAuth login, a full-scope key otherwise). --scopes also lists every granted scope; --json emits the user + credentialType/canReadBalance/canSpend/scopes (scriptable).
civitai buzz [--json] Show your spendable Buzz balance (blue / green / yellow, plus a total). Needs the BuzzRead scope β€” a full-scope personal API key or civitai login --scopes generate; a default OAuth login token can't read it, and gets a clear message naming both fixes. --json emits {blue,green,yellow,total} (scriptable β€” handy for before/after diffing a dev:live spend).
civitai app list [--kind <k>] [--category <c>] [--sort <s>] [--limit <n>] [--cursor <c>] [--json] Discover published Apps in the store (GET /api/v1/apps) β€” filter-based discovery, not free-text search. Needs a credential (civitai login or CIVITAI_TOKEN): the endpoint keys the visible catalog off your identity, so this is not one of the anonymous reads. Cursor-paged. See Browse the App store.
civitai app view <slug> [--json] Show one published App's store detail (GET /api/v1/apps/{slug}) β€” description, category, rating, gallery, live/external target. Needs a credential, same as app list. Reads the public store catalog, which is a different resource from your own deploy β€” a not-found here says nothing about <slug>.civit.ai. See Browse the App store.
civitai app create [name] [dir] [--template static|page-vite|page-money] [--dir <path>] [--name <display>] [--slug <slug>] [--yes] The friendly happy path. Scaffold a ready-to-build App, defaulting to the batteries-included page-money SDK template (default dir ./<slug>). --slug sets the blockId explicitly instead of deriving it from the name β€” needed when derivation refuses the name (see The blockId). -y/--yes is the non-interactive form: it never prompts, taking flags and defaults instead, and fails if a name is missing.
civitai app init [name] [dir] [--yes] [...] Same scaffolder as create with a no-build static default (back-compat alias); same --yes.
civitai app dev-token <slug> [--env] [--spend] [--budget <n>] Mint a short-lived (~4h) dev block token for npm run dev:live. --spend must be asked for explicitly to request real-Buzz spend β€” without it the CLI filters ai:write:budgeted out of the mint request. --env prints a paste-ready VITE_LIVE_BLOCK_TOKEN=<token>. See Local dev loop.
civitai app dev-tunnel [blockId] [--block <id>] [--port <n>] [--local-host <host>] [--tunnel-endpoint <h:p>] [--idle-timeout <d>] [--ready-timeout <d>] [--no-wait] (Pre-GA / invite-gated) Preview your local dev server inside the real Civitai host at civitai.com/apps/dev/<blockId> β€” a prod-fidelity inner-dev-loop. Pre-flights whether the host can actually embed your dev server and warns (never fatally) when it cannot. See Preview in the real host.
civitai app validate [dir] [--strict] [--json] Best-effort local pre-check of block.manifest.json; emits non-fatal warnings (--strict fails on them). --json emits the structured result (ok, plus errors/warnings each with field/message β€” field is always present and never null) for scriptable parsing β€” still exits non-zero on failure. πŸ”΄ BREAKING: a [dir] that does not exist, or is not a directory, is now a usage error β€” exit 2 with no JSON object on stdout, where it used to print {"ok": false, …} and exit 1. See Validate fidelity and The --json result shape.
civitai app submit [dir] [--yes] [--package-only] [--out f.zip] [--skip-validate] Validate + package the source tree + upload it with your stored token (or, with no token, write the bundle + print next steps). A submit that would really upload asks for confirmation, and in a non-interactive shell it refuses without --yes β€” civitai app submit --yes is the CI form. (The refusal is reached only when there is a token to upload with: --package-only, and the no-token fallback that just writes the .zip, never submit and so never ask.)
civitai app pull [dir] --app <slug|appBlockId> Clone (or sync) the canonical git repository behind one of your approved Apps β€” the read side of git authoring. ⚠ The clone URL embeds your access token, and a fresh clone persists it into .git/config. See Pull your app's repository.
civitai app listing status|set-icon <file>|set-cover <file>|add-screenshot <file>|rm-screenshot <id>|reorder <id...> Attach the store-listing media your App needs before it can be published β€” an icon and a cover are mandatory (screenshots are optional, up to 8). listing status prints what is attached vs. what the publish floor still requires. The CLI checks format + byte size locally; dimensions and aspect ratio are checked by the platform at attach. See After you submit and Listing media requirements.
civitai app status [blockId] [--id <pubreq>] [--json] Check the review/deploy status of your own submissions. No arg lists them all; a blockId (app slug) or --id shows one in detail (rejection reason if rejected, live URL once deployed). See Submission status.
civitai app metrics <slug> [--from <d>] [--to <d>] [--json] Owner-only analytics for one of your Apps β€” installs, runs + Buzz spent, Buzz purchased, and API engagement. Always prints the window the server served (it defaults to 30 days and clamps to 366), so a zero is never ambiguous. Needs a personal API key (an OAuth login is refused). See App metrics.
civitai app withdraw [pubreq-id] [--id <pubreq>] Withdraw your own pending submission (the pubreq_… id from civitai app status). Frees the slug so a fresh civitai app submit can replace it. Idempotent; only a pending request can be withdrawn. See Submission status.
civitai generate "<prompt>" [--negative-prompt <p>] [--quantity <n>] [--aspect-ratio <r>] [--checkpoint <version-id>] [--lora <version-id>[:strength]] [--image <path-or-url>] [--ecosystem <key>] [--input <file>] [--print-input] [--dry-run] [--json] [--max-cost <buzz>] [--fail-on-substitution] [--yes] [--no-wait] [--timeout <dur>] [--out-dir <dir>] [--no-download] [--force] [--external-id <key>] Generate images from a text prompt β€” this SPENDS REAL BUZZ. Prices the job with the server's estimator, shows the cost + your balance, asks before spending, submits, then waits and downloads the results. --dry-run prices it and exits without submitting; --max-cost is an estimate check, not a spending cap. Needs the AI Services scopes β€” civitai login --scopes generate or a full-scope personal API key; a default OAuth login is refused. See Generate for the wait/download flags, image-to-image, raw graphs, and silent model substitution.
civitai workflows list [--limit <n>] [--cursor <c>] [--tag <t>] [--json] List the generation workflows you have submitted, newest first β€” status, when, cost, and deliverable/total outputs. Cursor-paged: the next cursor is printed on stdout when more results exist. Reading spends nothing. See Generate.
civitai workflows get <workflow-id> [--json] Look up one generation workflow β€” status, steps and outputs. This is how you re-attach after --no-wait, a --timeout expiry or a Ctrl-C. Outputs that are blocked, unavailable or hidden are listed with the reason rather than omitted. Output URLs are presigned and expire; re-run for fresh links. Reading spends nothing. See Generate.
civitai workflows cancel <workflow-id> [--yes] [--json] Stop a running generation. πŸ”΄ This does not undo the charge β€” a mid-run cancel bills the cost already accrued. Cancel because you no longer want the output, never to save money. Asks for confirmation (default no); --yes skips the prompt and a non-TTY without it refuses. See Generate.
civitai upgrade [--force] Self-update this binary in place β€” resolve the latest GitHub release, verify its SHA-256 against checksums.txt, and replace the running executable. A Homebrew install delegates to brew upgrade instead; --force reinstalls anyway (and self-replaces a Homebrew install). See Upgrading.
civitai version Print version / commit / build date.
civitai completion [shell] Generate a shell-completion script.

Run civitai help, civitai app --help, or civitai <command> --help for the full details and examples.

The blockId

The blockId is your app's permanent public identity: the hostname it will be served at once approved (https://<blockId>.civit.ai/) and the argument every later command takes (app status, app metrics, app listing, app dev-token, app dev-tunnel). It cannot be renamed afterwards. app create / app init echo the one they chose, so it is on screen before you commit anything.

By default it is derived from the name: "My Cool Block" β†’ my-cool-block. Pass --slug <slug> to choose it yourself β€” it bypasses derivation entirely, so name, blockId and directory are three fully independent axes.

Breaking change. Derivation used to lowercase the name and replace every run of non-[a-z0-9] with a hyphen, which silently dropped characters: civitai app create "CafΓ© App" minted the blockId caf-app, and "ÜberApp" minted berapp β€” a different permanent public id than the author typed, with no warning and exit 0. Derivation now refuses and names the offending characters, exiting 2 and asking for --slug. If you have a script passing a non-ASCII name, it must now pass --slug <slug>. The old output was wrong, so the break is the point β€” but it is a break.

What derivation refuses is letters, digits and marks above ASCII that the slug alphabet cannot carry. Three things still derive rather than refuse, and they are deliberate:

input blockId why
"Rocket πŸš€ App" rocket-app Symbols, emoji and non-ASCII punctuation are separators β€” that is what makes "Widget β€” Pro" β†’ widget-pro right. An emoji has no lossless ASCII form either, so refusing would only trade a silent drop for a dead end. Tracked as #272.
"Δ°stanbul App" istanbul-app Exactly two runes above ASCII lowercase into ASCII β€” Δ° (U+0130) and K (U+212A). Lowercasing is what decides whether a character survives, so these transliterate for free.
"My Cool___Block" my-cool-block ASCII is exempt by construction β€” every derivation that worked before still produces the byte-identical blockId.

A name that is not valid UTF-8 is refused outright (it used to lose the bad bytes from the blockId and write them into block.manifest.json).

Templates
  • static β€” a no-build page app (index.html + a tiny app.js, block.manifest.json with page:{}, no build step).
  • page-vite β€” a Vite + React page app with config-as-code build fields (buildCommand: "npm run build" + outputDir: "dist").
  • page-money β€” a Vite + React + TypeScript full-page (W10) money-path app wired to the published App SDK (@civitai/blocks-react + @civitai/app-sdk): prompt β†’ estimate β†’ lazy consent β†’ submit β†’ poll β†’ real Buzz spend, via useBuzzWorkflow / useRequestConsent / useBlockResize (never raw postMessage). Ships a dev:harness mock host, .env.* allowed parent-origin config, and a unit-test stub. Run npm run dev:harness (plain npm run dev renders blank without a host).

Every template also scaffolds an assets/ directory holding a README of the store-listing media requirements β€” and no images, so the set-icon / set-cover step fails loudly until you supply real artwork. See Listing media requirements.

The host handshake (BLOCK_READY)

Every template declares a page surface, and the host will not reveal a page app until the app posts BLOCK_READY β€” that handler is the only transition into the host's ready state. An app that never sends it is replaced by a visible failure card once the host's bounded retries run out, even though the app itself renders perfectly. Nothing you can run locally reproduces that.

page-money gets the handshake for free: @civitai/blocks-react's iframe transport acks internally, which is why the SDK templates never touch raw postMessage. The two SDK-free templates (static, page-vite) therefore ship a small vendored emitter, civitai-host.js, loaded from the entry point. Leave it in place β€” and if you are retrofitting it into an older app, note that the file has to be referenced as well as copied: static loads it with <script src="./civitai-host.js"></script> in index.html, page-vite with import './civitai-host.js'; as the first line of src/main.jsx. civitai app validate checks that reference where it can resolve your entry point, and tells you when it can't.

⚠️ If you adopt @civitai/blocks-react, delete civitai-host.js in the same change. This is the one situation where removing it is correct, and running both is worse than running neither: whichever handshake answers the host's first BLOCK_INIT cancels the host's retry loop and its readiness timeout. If the vendored emitter wins that race, the SDK transport can be left never having seen an init β€” its waitForInit rejects after 10s and the host sits "ready", showing an app that never started, with no retry and no error card.

Two rules it encodes, which apply to every message you add afterwards:

  • The envelope is { type, payload }. The host dispatches event.data.payload to its subscribers, so fields put at the top level ({ type: 'X', height: 0 }) arrive as payload: undefined.
  • Answer, don't announce. The ack goes out in response to the host's BLOCK_INIT, addressed at the origin that init arrived from rather than broadcast to '*'. It is also why nothing is posted when you preview locally: there is no host to send BLOCK_INIT, so the emitter stays silent by design.

πŸ”’ The emitter checks the sender window, not the sender's identity. It answers window.parent β€” whoever framed you β€” which is sound for this one message because the ack carries no data. It is not sufficient for anything you add next. The moment you handle an inbound message carrying a token, a viewer, storage or a result, check event.origin against an allowlist of origins you trust, or any page that frames your app can feed it whatever it likes. The emitter deliberately does not vendor that allowlist β€” the real list (production, preview subdomains, dev tunnels) is platform state that moves without notice, and @civitai/blocks-react already maintains it from VITE_BLOCK_ALLOWED_PARENT_ORIGINS. Adopt the SDK before you handle data.

RESIZE_IFRAME is not part of a page app's protocol: the host renders a page block full-viewport, so it does not size to content and ignores the message. (useBlockResize is surface-agnostic and page-money still calls it β€” on a page surface it is simply a no-op, which is why the SDK templates can share component code across surfaces.) The pre-#206 templates demoed a raw postMessage of RESIZE_IFRAME, so a project scaffolded before that fix still carries dead code you can delete. (This CLI's own CI fails if a shipped template ever reintroduces it; there is no author-facing command that scans your project for it β€” civitai app validate checks the manifest and the handshake, not this.)

Local dev loop (harness: mock vs live)

A scaffolded App is a sandboxed iframe, and locally there is no host to send BLOCK_INIT β€” so npm run dev shows you your own UI and nothing of the protocol. The page-money template ships a dev harness (the SDK's @civitai/blocks-react/testing hosts) to close that gap, with two modes:

Command Mode What it does
npm run dev:harness mock (default) Mounts the SDK mock host β€” synthetic replies, no real Buzz, no compute, no network. Safe to spam; drive money/error/insufficient-Buzz UX via on-screen scenarios or ? URL params. Start here.
npm run dev:live live Mounts the SDK live host (createLiveHost) β€” forwards the App protocol to the real Civitai backend with a pasted dev token (Bearer). Spends REAL Buzz / real compute.

⚠️ dev:live works on a pending (un-approved) app. The dev-token mint (POST /api/v1/blocks/dev-token) accepts a pending slug β€” right after a successful civitai app submit (status pending) it returns 200 with appId: pending-pubreq_… and dev:live mounts the live host against the pending app. For real generation you must mint with a credential carrying AI Services β€” a full-scope personal API key, or an OAuth login that opted in via civitai login --scopes generate. A default OAuth login (civitai login, no --scopes) mints read-only (user:read:self) and cannot spend. Use civitai buzz / civitai whoami to confirm your credential can spend before a live run, and pass --spend to dev-token to request the spend scope explicitly β€” without it the CLI filters ai:write:budgeted out of the request, so a spend-capable credential still mints a token that will not generate.

Live mode needs a short-lived dev block token. Mint it with civitai app dev-token (the CLI handles the invite-gated POST /api/v1/blocks/dev-token call with your stored credential β€” no hand-rolled curl) and paste it into .env.development.local as VITE_LIVE_BLOCK_TOKEN=:

# From your scaffolded project dir (reads scopes from block.manifest.json):
civitai app dev-token my-block --env >> .env.development.local
npm run dev:live

dev-token reads the scopes to request from your local block.manifest.json, so it works on a slug you have never submitted. .env.development* is never committed (submit excludes it) and the token is short-lived (~4h) β€” re-run dev-token when it expires. For real generation mint with a spend-capable credential (full-scope personal API key or civitai login --scopes generate) and add --spend: the CLI never asks for budgeted spend implicitly β€” even when your manifest declares it (the scaffolded money app does) β€” so without the flag the mint filters ai:write:budgeted out and dev:live refuses to generate with block lacks ai:write:budgeted scope. --budget <n> sets the token's per-generation Buzz budget (1–250; omit it and the server picks one β€” 50 for an unsubmitted app). A default OAuth login mints a read-only token either way (the command warns you at mint time). With no token, dev:live fails safe (renders a notice, never spends). Live v1 covers the money path (estimate/submit/poll/cancel); pickers, checkpoint-set, App-Storage KV, and in-band Buzz purchase are mock-only.

Under the hood (the scaffold wires this β€” you don't configure it): dev:live routes the live host's backend calls through the vite dev proxy (server.proxy['/api']), not straight to civitai.com: createLiveHost fetches /api/... SAME-ORIGIN against the dev server (localhost:5186), and vite proxies that server-side to civitai with the Origin header rewritten to an allowlisted host. This is load-bearing β€” a direct cross-origin fetch from localhost is both blocked by CORS preflight and rejected by civitai's tRPC origin gate. The same-origin proxy + Origin rewrite fixes both. VITE_LIVE_HOST_ORIGIN overrides the proxy target (default https://civitai.com).

Which credential can spend? Spending Buzz β€” a real dev:live generation in your app, or a civitai generate run from the terminal β€” needs the AI Services scope. Two credentials carry it; the default OAuth login deliberately does not:

Credential Can spend Buzz? (dev:live, civitai generate) How to get it
Personal API key (full scope) βœ… Yes β€” estimate β†’ submit β†’ generation β†’ real Buzz create it in the web UI at civitai.com/user/account, then civitai login --token <key> (a personal key carries AI Services)
civitai login --scopes generate (OAuth, opt-in) βœ… Yes β€” additive on top of the default, so it also keeps submit + dev-tunnel civitai login --scopes generate (the civitai-cli client's allowedScopes on civitai.com includes the generate bits β€” see Submit & auth)
civitai login (OAuth, default) ❌ No β€” viewer + catalog + app storage only the default scope set omits AI Services, so the server strips the spend scope β€” fine for read/identity dev:live, not for generation

This is the single most common blocker for civitai generate: a default OAuth login looks perfectly valid, and the refusal is a scope problem, not a login problem β€” re-running plain civitai login will not fix it, but civitai login --scopes generate (or a full-scope personal key) will. civitai whoami shows the capability as Spend Buzz (AI Services) and names the fix for whichever credential you have.

You can't mint a personal key over OAuth or the CLI (apiKey.add returns 403 without a full-scope session) β€” create it in the web UI. The dev token always grants user:read:self, so your viewer resolves on either path. For the scope mechanics behind this, see Submit & auth.

Env vars (VITE_BLOCK_ALLOWED_PARENT_ORIGINS, VITE_HARNESS_MODE, VITE_LIVE_BLOCK_TOKEN, …) and the scenario knobs are documented in depth in the scaffolded project's own README.md and .env.example β€” see internal/scaffold/templates/page-money/README.md.tmpl.

Preview in the real host (app dev-tunnel)

(Pre-GA / invite-gated.) Access is gated behind an Apps-author invite and a server kill-switch flag, so if you are not enrolled the mint reports "not available" β€” ask to be added to the cohort.

The harness is a mock of the host. civitai app dev-tunnel is the other end of that trade: it previews your local dev server inside the real Civitai host at civitai.com/apps/dev/<blockId>, so you see prod chrome, prod sandbox and the prod handshake against the code in your editor.

npm run dev:tunnel                 # start your dev server first
civitai app dev-tunnel my-block    # then open the tunnel

It mints an ephemeral in-memory ssh keypair, opens a reverse tunnel from your dev port to the Civitai tunnel endpoint (sish.civitai.com:2224, live), prints the URL to open, and tears everything down on Ctrl-C or an idle timeout. Publishing the tunnel host through external-dns + Cloudflare usually takes 1–3 min (occasionally longer); the command waits for it and prints the elapsed time.

Embeddability preflight. Before minting, the command checks whether the host can actually embed your dev server. The host iframes it sandboxed, at an opaque null origin, so a dev server that is missing Access-Control-Allow-Origin: *, missing the .civit.ai entry in allowedHosts, or sending a framing header that excludes civitai.com loads as a blank iframe with no error anywhere β€” the worst kind of failure to debug. Those are printed as warnings, never fatal, and deliberately twice: once the moment the checks run, so you still get them if you Ctrl-C the DNS wait, and again just above the URL, with the vite.config.ts fix. Apps scaffolded by civitai app create (the page-money template) already satisfy all of it.

Flags. The defaults match what the scaffold's npm run dev:tunnel binds, so most authors pass none of these:

Flag Default What it is for
--block <id> the blockId in block.manifest.json in the CWD The app to tunnel, if you are not standing in its project (it is also the positional argument).
--port <n> 5186 The local dev-server port. Matches the scaffold's dev:tunnel script.
--local-host <host> localhost The host your dev server is bound to. Change this when the dev server is not on the CLI's own loopback β€” inside a container or pod (--local-host 10.42.0.100), a VM, or bound to one specific interface. A dev server that is reachable in your browser but not from the CLI's localhost is exactly what this flag is for.
--no-wait off Print the public URL immediately instead of waiting for it to start serving. It may 404/NXDOMAIN for a few minutes while DNS and routing propagate.
--ready-timeout <d> 0 (wait indefinitely) Cap the readiness wait. On expiry the command warns and prints the URL anyway rather than failing.
--idle-timeout <d> 30m Tear the tunnel down after this much inactivity.
--tunnel-endpoint <host:port> sish.civitai.com:2224 The sish SSH endpoint to dial. Also settable with CIVITAI_DEV_TUNNEL_ENDPOINT; the flag wins.

Publishing DNS takes 1–3 minutes, sometimes longer. That wait is normal and the command reports elapsed time while it happens β€” it is not a hang. If you Ctrl-C out of it you also lose the embeddability warnings, which is why they are printed once before the wait as well as again after it.

Examples

Two real example manifests live under examples/ (copied from the civitai-block-* dogfood apps). Read them for manifest shape β€” between them they cover the required fields, $schema wiring, the page/iframe blocks, and scope declarations with justifications:

The values are those apps' own choices, not recommendations. In particular don't copy buzz-generator's page.buzzBudgetPerGen β€” it is a safety ceiling against a malicious or compromised app, not an estimate of one run, so size your own from the field's description in the canonical schema (notepad doesn't take the budgeted scope, so it has no budget at all).

Both validate clean (examples_test.go asserts this so the claim stays true) β€” schema conformance only, which says nothing about whether a value is well-sized.

Browse the public API

Beyond authoring Apps, the CLI is a thin client for Civitai's public read REST API (GET /api/v1/**). These subcommands work anonymously β€” no login needed, because the data is public β€” but when you're logged in your stored token is sent automatically (pass --anon to force a no-auth request). Every command also takes --json to print the raw API JSON response for scripting.

Command What it does Notable flags
civitai models search Search models (GET /api/v1/models) --query, --tag, --username, --type, --base-model (repeatable), --sort, --period, --nsfw; paging --limit (≀100), --page, --cursor
civitai models get <id> Get one model by id --json, --anon
civitai model-versions get <id> Get a model version by id (alias mv) --json, --anon
civitai model-versions by-hash <hash> Look up a model version by file hash (AutoV2, SHA256, …) --json, --anon
civitai download <version-id> Download a model version's file(s) --model, --file, --all, --out, --out-dir, --layout, --root, --for-base, --no-verify, --force, --anon
civitai images get <id> Get one image by id (GET /api/v1/images?imageId=<id>) --json, --anon
civitai images search Search images (GET /api/v1/images) --model-id, --model-version-id, --post-id, --username, --base-model (repeatable), --type (image/video/audio), --sort, --period, --nsfw, --meta (include generation metadata); paging --limit (≀200), --page, --cursor
civitai tags search Search model tags --query; paging --limit (≀200), --page
civitai creators search Search creators --query; paging --limit (≀200), --page
civitai users get <username-or-id> Look up a user via public search (a number = exact id; a name = exact-username match, else it lists close matches) --json, --anon
civitai articles search Search articles (GET /api/v1/articles) --query, --tags, --username, --sort, --nsfw; paging --limit (≀100), --cursor
civitai articles get <id> Get one article by id (--content renders the article body as readable text/markdown) --content, --json, --anon
civitai collections search Search public collections (GET /api/v1/collections) --query, --sort, --nsfw; paging --limit (≀100), --cursor
civitai collections get <id> Get one collection by id --json, --anon

Pagination. List commands print a compact footer with the next-page hint. models/images support both shallow --page and deep --cursor paging (the API caps page*limit at 1000 and 429s beyond it β€” prefer --cursor for deep paging); articles/collections are cursor-only (keyset feed β€” no --page); tags/creators are --page-only. Each endpoint caps --limit (models/articles/collections 100; images/tags/creators 200).

civitai models search --query "pony" --limit 5
civitai models get 4384
civitai model-versions by-hash 5D8D26E2A6
civitai articles get 32680
civitai articles get 32680 --content   # render the article body (the guide) as readable text/markdown
civitai images search --model-id 4384 --sort "Most Reactions" --json   # raw JSON for scripting

Filtering by base model. --base-model is repeatable and maps to the REST baseModels filter (an OR across the values). It's the key discovery filter for things --type can't separate β€” e.g. video checkpoints all share --type Checkpoint and are distinguished only by base model. It works on both models search and images search:

civitai models search --type Checkpoint --base-model "Wan Video 2.2 T2V-A14B"
civitai models search --base-model Pony --base-model Illustrious --limit 20
# images too β€” find recent-popular images generated with a given base model:
civitai images search --base-model "Krea 2" --sort "Most Reactions" --period Week
civitai images search --type video --sort "Most Reactions"   # videos only

Generation metadata (--meta). By default the image list is a compact table without generation data (matching the API, which omits meta unless asked). Add --meta to include each image's prompt, sampler, cfg, steps, seed, and model β€” rendered as an indented detail block per image (the table can't hold a prompt). Images whose uploader chose to hide their generation data show meta: (hidden by uploader). With --json, --meta adds the raw meta object to each item.

civitai images search --nsfw --sort "Most Reactions" --period Month --meta
civitai images search --model-version-id 128713 --meta --json | jq '.items[].meta'

The human table includes a BASE MODEL column (the base model each image was generated with, when the API reports one; - when it doesn't), so you can see the ecosystem at a glance without dropping to --json.

--sort is ignored with --model-id. The REST API returns images for a given modelId in its own default order regardless of sort, so images search --model-id <id> --sort … prints a one-line note on stderr and the results are NOT re-sorted. (--model-version-id is unaffected β€” it honours --sort.)

Non-weights file marker. In the human (non---json) output of models get and model-versions get, a version whose primary file is not model weights (type != "Model") is tagged with its actual file type β€” e.g. [Archive] (a "Workflows" model's downloadable deliverable), [Training Data], or [Other] β€” so you can see at a glance that the version's file isn't weights. It's purely informational: any file type still downloads. --json output is an unchanged raw passthrough.

Download model files

civitai download fetches the file(s) of a model version. Identify the version deterministically by its numeric version id, or resolve a model's default (first) published version with --model:

civitai download 691639                       # the version's primary file β†’ ./<server-name>
civitai download --model 4384                 # resolve model 4384's default version, then download its primary file
civitai download --model 4384 --dry-run       # print the plan (files, sizes, hashes, targets) β€” download nothing
civitai download 691639 --out ./flux_dev.safetensors
civitai download 290640 --file vae --out-dir ./models   # pick a file by name; write into a dir
civitai download 691639 --file 1234567                  # pick one of two same-named files by its file id
civitai download 290640 --all --out-dir ./models        # every file in the version
civitai download 290640 --all --layout comfyui --root ~/ComfyUI   # route each file to its type folder
civitai download 691639 --layout a1111 --for-base "SDXL 1.0"      # A1111 layout + base-model compat warning

Downloads require authentication. Every model-file download needs a token β€” even a small public embedding 401s anonymously. Run civitai login first. The read/search commands work anonymously; downloads do not. --anon is meaningful for the read commands, not for download.

Behavior:

  • Identifier β€” exactly one of the positional id or --model <model-id> is required. The positional is normally a model-version id, but because models search / models get print model ids, handing one over just works: the CLI notices it is a model id and downloads that model's default version, printing a note: <id> is a model id β€” downloading its default version <v> line. When a number is both a valid model id and a valid version id (common for low/mid numbers), the CLI stops rather than guess, naming both interpretations β€” re-run with --model <id> (that model's default version), --version <id> (that version as-is), or --yes to take the version interpretation and have it echoed back. --version names a version id explicitly and skips the stop entirely.

  • --model resolves the default version β€” the model's default (first published) version; its primary file is downloaded regardless of file type. Any model type works, including a type: Workflows model whose deliverable is a downloadable Archive.

  • --dry-run β€” resolve the version + selected file(s) and print the plan (each file's name, size, SHA256, resolved target path, and whether authentication will be required) then exit 0, transferring nothing and creating no file (not even a .part). Works with --file, --all, --model, --out, --out-dir, and --layout/--root (the plan shows the routed target paths).

  • File selection β€” defaults to the version's primary file. --file selects one file by numeric file id (the version's files[].id) or by name (exact, else a unique case-insensitive substring; ambiguous/none errors and lists the candidate files with their ids). --all downloads every file.

  • Same-named files (no silent overwrite) β€” a version can ship two files that share a name (e.g. Flux Dev's fp16 and fp8, both flux_dev.safetensors). Selecting that shared name with --file is ambiguous and errors, listing both files with their ids β€” pass the numeric id to pick exactly one (--file 1234567; the id is shown by --dry-run and in the error). --all refuses to run when two selected files would resolve to the same on-disk path (which would silently clobber one) β€” it fails before transferring anything, lists the colliding files with their ids/sizes, and tells you to pick one with --file <id> (or write them to separate paths). No download ever silently overwrites another.

  • Output β€” --out <path> sets an exact target path (single file only). --out-dir <dir> writes server-named files into a directory (works with --all). Parent directories are created as needed. Default is the server-provided filename in the current directory.

  • Type-aware folder routing (--layout) β€” --layout <a1111|comfyui> writes each file into the correct subfolder for that app, keyed by the file/model type, under --root <dir> (default .). This fixes the footgun where --all --out-dir X dumps a bundled VAE into the checkpoint folder and pollutes the model dropdown: with --layout, the checkpoint lands in the checkpoints folder and the VAE in the VAE folder. --layout is mutually exclusive with --out/--out-dir; --root only applies with --layout. An unmapped type (Poses, Wildcards, Archive, …) is written to --root with a stderr note rather than silently misplaced. The routed folder maps:

    Civitai type A1111 / Forge ComfyUI
    Checkpoint models/Stable-diffusion models/checkpoints
    VAE (standalone or bundled) models/VAE models/vae
    LORA / LoCon / DoRA models/Lora models/loras
    TextualInversion (embedding) embeddings models/embeddings
    Hypernetwork models/hypernetworks models/hypernetworks
    Controlnet models/ControlNet models/controlnet
    Upscaler models/ESRGAN models/upscale_models

    (Sources: the AUTOMATIC1111 wiki + the sd-webui-controlnet models/ControlNet default; the ComfyUI models docs.)

  • Mis-file warning (without --layout) β€” when --all would place files of differing types into one directory (the mis-file footgun), the CLI prints a one-line stderr warning naming the off-type file(s) and suggesting --layout. It's a warning, not an error; a single-type download stays quiet.

  • Base model + compatibility (--for-base) β€” the version's base model is always shown in the plan/output. --for-base "<baseModel>" warns on stderr when the version's base model is in a confidently different family than your target (e.g. an SD 1.5 embedding like EasyNegative downloaded for an SDXL 1.0 model β†’ won't work; the wrong VAE β†’ black images). The check is conservative β€” it groups the common bases into architecture families (SD1.x, SD2.x, the SDXL family [SDXL/Pony/Illustrious/NoobAI, treated loosely], SD3, Flux, video, …) and only warns on an architecture-level mismatch, never on near-neighbours (Pony vs Illustrious) or unclassifiable bases.

  • Streaming + atomicity β€” the body streams to <target>.part and is renamed into place only on success, so an interrupted run never leaves a truncated final file. Large files (10+ GB) are never buffered in memory. TTY-aware progress is printed to stderr. The Civitai download URL 302-redirects to signed storage; the CLI follows it.

  • Auth β€” your stored login token (civitai login) or CIVITAI_TOKEN is used automatically; Civitai requires a token to download any model file, even public ones, so an anonymous download gets an actionable 401 (401 β†’ run civitai login; 403 β†’ the file is gated for your account). --anon forces no token.

  • Transient-failure retry (reads) β€” the read endpoints (search / model / version / images / tags / creators / users / articles / collections) retry a transient 502/503/504 or network error a few times with exponential backoff (with jitter), noting each retry on stderr. A 429 is retried only when it carries a Retry-After header (a genuine throttle, honored up to a cap); a 429 without Retry-After is Civitai's deterministic deep-paging limit and is surfaced immediately with the hint to use --cursor instead of --page. The download stream is not retried mid-transfer.

  • Integrity (default on) β€” the streamed bytes are verified against the file's SHA256; a mismatch deletes the .part and fails. --no-verify skips it; a file with no published SHA256 downloads with a warning (not a hard failure). Note that SHA256 verifies integrity (the bytes match what the API advertised), not authenticity β€” it proves the download wasn't corrupted or truncated in transit, but a compromised source that advertises a matching hash for malicious bytes cannot be detected by the hash alone. Only download models from creators you trust.

  • Pickle/archive safety note β€” when a downloaded file has a pickle/executable extension (.ckpt, .pt, .pth, .bin, .pickle, .pkl) or an archive extension (.zip, .tar, .tar.gz, .tgz, .rar, .7z), the CLI prints a one-line stderr note: these formats can execute arbitrary code when loaded by ComfyUI/A1111/torch.load, and they land in folders those apps auto-scan. safetensors and image files are inert and get no note. The note is informational β€” it never blocks the download.

  • ControlNet preprocessor note β€” when the parent model is a ControlNet, the CLI prints a one-line stderr note: a ControlNet model needs a matching preprocessor/annotator (e.g. the ComfyUI comfyui_controlnet_aux custom node β€” OpenPose/Canny/Depth) to derive the control image from your input, and that preprocessor is a separate install, not hosted on Civitai. The note is informational β€” it never blocks the download.

  • Idempotency β€” an already-present target (that verifies, or with --no-verify) is skipped with a note; --force re-downloads.

  • Any file type downloads β€” the selected/primary file is downloaded whatever its type (Model weights, a type: Workflows model's Archive, training data, or other artifacts). The human models get / model-versions get output tags a non-weights primary file with its type (e.g. [Archive]) purely for information; it never blocks a download.

Scripting with --json

Every read subcommand takes --json, which prints the raw /api/v1/... REST response β€” a stable passthrough, not a CLI-invented shape. So the field schema is exactly the public Site API's; keep the REST field reference open (e.g. models, model-versions) rather than reverse-engineering fields with jq keys.

Two properties make the output safe to pipe:

  • --json stdout is pure JSON β€” nothing else is written to stdout, so ... --json | jq -e . always parses.
  • Errors go to stderr with a non-zero exit β€” a failed call writes the error to stderr, exits non-zero, and prints nothing to stdout, so jq never sees error prose. For example civitai model-versions get 999999999 --json exits 4 with Error: not found (404): Model not found on stderr and an empty stdout.

Both properties hold for civitai generate and civitai workflows … too, but their payloads are not Site API REST shapes β€” generation has no REST route, so those commands pass through the raw orchestrator reply. Read Generation --json before scripting against them.

Cursor pagination loop

For deep paging use --cursor (not --page β€” the API caps page*limit at 1000 and 429s beyond it). Read .metadata.nextCursor from each response and feed it back via --cursor; stop when it's absent/null:

export CIVITAI_NO_UPDATE_CHECK=1
cursor=""
while :; do
  page=$(civitai models search --type LORA --base-model Illustrious \
           --sort "Most Downloaded" --limit 5 ${cursor:+--cursor "$cursor"} --json) || break
  echo "$page" | jq -r '.items[].id'                 # do your work here
  cursor=$(echo "$page" | jq -r '.metadata.nextCursor // empty')
  [ -z "$cursor" ] && break                          # no more pages
done
Clean output for pipelines

The CLI runs a background check for a newer release and prints a nag to stderr. In scripts, silence it with CIVITAI_NO_UPDATE_CHECK=1 (env) or --no-update-check (flag). Either way stdout stays pure JSON β€” the nag never touches stdout β€” but suppressing it keeps stderr clean for logs.

Generation --json

civitai generate --dry-run --json, civitai workflows list --json and civitai workflows get <id> --json emit the raw orchestrator payload. Two caveats have bitten people, and neither shows up as an error:

  • Output URLs are presigned and EXPIRE. The links in a workflow payload are short-lived signatures, not durable addresses. A pipeline that stores them and fetches later gets a 401/403 from the storage host that no credential can fix β€” re-run civitai workflows get <id> for fresh links instead of caching the old ones. (Fetch them with no Authorization header; they are already authorized, and the CLI deliberately attaches nothing to them.)

  • --json still exits 0 when the server reports the resources are unavailable. --dry-run --json prints the estimate and exits 0 even when the payload says "ready": false. A human --dry-run prints a warning and this CLI refuses to submit in that state, but a script reading only the exit code sees success. Branch on the field, exactly as app metrics requires branching on notOwned β€” and note the shape below, which fails closed:

    q=$(civitai generate "a cat" --checkpoint 128713 --dry-run --json) || exit $?
case "$(printf '%s' "$q" | jq -r 'if has("ready") then .ready else "absent" end')" in
  false)   echo "resources unavailable" >&2; exit 1 ;;   # decisive: do not submit
  true)    ;;                                            # NOT a green light β€” see below
  *)       echo "no readable .ready field" >&2; exit 1 ;; # absent, null, or jq failed
esac
printf '%s' "$q" | jq -r .cost.total

    The * arm is the point. An earlier version of this snippet tested [ … = "false" ] && exit 1, which exits 0 when the key is absent or jq fails β€” it read "we could not ask" as "we asked and it was fine", the fabricated-zero mistake app metrics documents for views.unavailable.

    πŸ”΄ ready is one-directional, and the human label says so. It reports only that the resources this job needs are currently available β€” the server computes it as "every job's queue position reports support: available", and a job carrying no queue position at all is skipped, leaving the flag true. It is not a moderation verdict and not a prediction that the job produces an image; --dry-run therefore prints it as Resources ready, not "Generatable". A run reporting ready: true can still be charged and return nothing β€” measured: 8 submits across 3 checkpoints that all quoted ready: true produced 0 outputs. So gate on the FALSE direction, as above, and never treat true as a success predicate. The only thing that settles whether a job produced output is the finished workflow (civitai workflows get <id>), and civitai generate exits non-zero when it waited and got no deliverable output.

    What ready: false gets you is a LOCAL refusal, and that is all this repo can evidence. civitai generate reads the flag and refuses to submit; no server-side enforcement of it has been found β€” support !== 'available' appears once in the whatIf reply builder and nowhere on the submit path, and the checkpoints that failed in the measurement above surfaced as HTTP 400s rather than as ready: false. Treat it as this CLI's own pre-flight, not as a promise about what the server would have done.

    Cost keys (cost.factors, cost.fixed) are server-owned and passed through verbatim, so treat them as an open map rather than a fixed set.

Gotchas
  • SHA256 is UPPER-case in the API/--json (e.g. 42BA94DF20CC0F4E6DF46E3C294587A2F8CF133BF0134185884EE1C9C5E108C4), while sha256sum emits lowercase. Case-fold before comparing if you roll your own verify (civitai download's built-in check is already case-insensitive): [ "$(echo "$api_sha" | tr A-Z a-z)" = "$(sha256sum file | cut -d' ' -f1)" ].
  • models search already embeds .modelVersions[] β€” each item carries its full versions, including files[].hashes.SHA256 and trainedWords. If you're iterating search results you usually don't need a follow-up model-versions get per version.
  • Creator + model-level download counts live only in the search response. model-versions get <id> returns a version, whose .model is just {name, type, nsfw, poi} β€” no creator, no model stats.downloadCount. If you started from a version and need those, fetch them from models search / models get and join on the model id (.modelId on the version).
Worked example β€” top LoRAs for a base model, then plan a download

Search β†’ pick versions with jq β†’ hand each version id to download with app folder routing. --dry-run prints the plan (files, sizes, hashes, target paths) without transferring, so this snippet is safe to copy-paste:

export CIVITAI_NO_UPDATE_CHECK=1
civitai models search --type LORA --base-model Illustrious \
    --sort "Most Downloaded" --limit 3 --json |
  jq -r '.items[].modelVersions[0].id' |
  while read -r vid; do
    civitai download "$vid" --layout comfyui --root ~/ComfyUI --dry-run
  done

Drop --dry-run (and civitai login first) to actually fetch the files β€” --layout comfyui routes each into its ComfyUI type folder.

Validate fidelity

civitai app validate is a best-effort LOCAL mirror of the platform's approve-time validator (BlockManifestValidator). The server is the source of truth at review time β€” passing validate locally is a strong pre-check, not a guarantee of approval.

It checks block.manifest.json against a vendored JSON Schema (schema/app-block.manifest.schema.json, syntactic shape) plus the ported semantic rules the server runs (sandbox trust-tier allowlist, page β‡’ iframe, required iframe sub-fields, the renderMode tier gate, targets[].slotId registry membership) and structural project checks. A few checks are necessarily approximate locally (the slot registry is vendored; per-app origin-binding/scope checks the CLI can't see are not reproduced).

It also mirrors one build-time rule, because the failure it prevents is otherwise an opaque server-side "build failed": your committed lockfile must match the package manager the platform derives from buildCommand. The platform build installs strictly from the lockfile β€” no registry re-resolve fallback β€” so "buildCommand": "pnpm run build" needs pnpm-lock.yaml, "yarn run build" needs yarn.lock, and npm run … / vite build / npx vite build / an omitted buildCommand all need package-lock.json. A mismatch or a missing lockfile is a hard validate error; an extra unused lockfile is a warning. Apps with no package.json are static β€” the platform never installs for them and they are never flagged.

The lockfile also has to be one, not merely exist. A package-lock.json must parse as JSON and declare a numeric "lockfileVersion" of 1 or more; a pnpm-lock.yaml or yarn.lock must be non-empty. That version rule is deliberately stricter than npm ci measures β€” npm states it as a precondition but will happily install from an otherwise-intact lockfile whose version key is 0, a string, null or absent β€” and validate keeps it because npm never writes those shapes, so a file carrying one was made by hand. An empty lockfile fails the platform build exactly like a missing one, so touch package-lock.json is not a fix β€” run the package manager and commit what it writes. If the lockfile cannot be read, or is implausibly large, validate says nothing rather than guessing: it never blocks a submit on a file it could not inspect.

Finally it emits one advisory about the host handshake: if your manifest declares a page surface and nothing your app loads posts BLOCK_READY, validate says so. That is the shape of an app scaffolded before the templates were fixed (#206) β€” it renders perfectly everywhere you can look locally and is replaced by a failure card in the real host. It is a warning, never an error: unlike the lockfile rule (where the platform build provably dies), this one infers runtime behaviour from static text and can be wrong, so it must not fail a correct project.

⚠️ Copying civitai-host.js in is only half the fix. A browser never fetches a file nothing references, so the emitter has to be loaded too β€” a <script src="./civitai-host.js"></script> in index.html, or an import './civitai-host.js'; at the top of the entry module index.html loads. Earlier releases of this check looked only for the text BLOCK_READY anywhere in your tree, so an unreferenced copy silenced it and a still-broken app validated clean. It now resolves what your index.html actually loads.

Four things follow:

  • It checks REACHABILITY where it can, and says when it can't. Starting at index.html it follows every <script src>, inline module and import it can resolve, and asks whether any of those files posts the message. When that resolution is complete you get a precise finding β€” including "you have an emitter, but nothing loads it". When it isn't β€” no index.html at your project root, a bundler alias (import '@/…'), a reference to a file that isn't there, an import chain deeper than it follows β€” it falls back to scanning your whole tree for the text, and the warning says so in as many words: "it did NOT check that the file is loaded". Read that sentence as it is written; in that mode, adding the emitter without referencing it will silence the warning and leave the app broken.
  • A dependency that acks ends the check β€” today that is @civitai/blocks-react, and nothing else. Its iframe transport acks internally and the literal never appears in your src/, so a page-money app is never flagged. This is an exact list, not the @civitai/ scope: @civitai/app-sdk is the server-side SDK and no runtime code in it posts BLOCK_READY, and @civitai/theme / @civitai/components are CSS. Depending on those does not give you the handshake, so it does not silence the check either.
  • It reads source only β€” never node_modules, never the conventional build directories (dist, build, out, …), and never a .md file: a README describing the handshake is not an implementation of it. Comments are stripped too, so a comment naming BLOCK_READY does not satisfy it. A src that is a symlink into a shared package is followed.
  • It stays quiet when it cannot see the whole project. An unreadable file, a file over 2 MiB, a very large tree, or a directory holding only a manifest all mean "we could not look" β€” reported as nothing, never as a finding. Likewise a project whose entry graph can't be resolved and which contains the literal somewhere: quiet, deliberately, because warning at a correct project is the more expensive mistake.

If it fires on a project you know is correct β€” your ack arrives from a bundled dependency, or from a file type this scan doesn't open β€” it is a false alarm, and it never blocks (exit 0) unless you pass --strict. What it proves stays narrow even at its strongest: that a file your index.html really loads mentions the message. It cannot prove the ack ever fires; only the real host can.

civitai app submit prints the same warnings before it uploads, and likewise does not block on them.

⚠️ If you already run civitai app validate --strict in CI, this advisory is new and can turn a previously-green project red β€” which is what --strict asks for. If it is a false alarm for your project, drop --strict or add the ack, and please open an issue: a warning at a correct project is a bug in the check, not something you should have to work around.

The durable fix is a server-side civitai app validate endpoint that calls the real BlockManifestValidator (the faithful contract), with this schema published as the syntactic half. See AGENTS.md for the full caveat and how the vendored schema + Go checks are kept in sync.

The --json result shape
{
  "ok": false,
  "dir": "./my-block",
  "errors":   [ { "field": "iframe.sandbox", "message": "…" } ],
  "warnings": [ { "field": "page.buzzBudgetPerGen", "message": "…" } ]
}

Two guarantees, both of which the previous release broke:

  • field is present on every finding, and is never null or empty. It used to be recovered by parsing the message text, which worked only for the JSON Schema errors β€” so every semantic finding (the sandbox rules, the justification rules, the money-path warnings: the ones a local pre-check exists for) arrived as "field": null, and grouping by field in CI silently dropped them. Findings now carry their field from where they are produced.
  • One notation: dotted paths. blockId, iframe.sandbox, scopes[1], targets[0].slotId, scopeJustifications.<scope> β€” the same way the human-readable messages, this README and the schema all name fields. Earlier releases mixed JSON Pointer (/blockId, /scopes/1) into --json while the text output used dotted; if you were matching on /-prefixed fields, update your scripts.

Two findings have no single manifest field, and say so explicitly rather than omitting the key:

field meaning
(root) the manifest document β€” it is missing, unparseable, or the schema reports a violation at the top level (e.g. missing property 'contentRating').
(project) repository state outside the manifest β€” the committed lockfile, or the source tree the BLOCK_READY advisory reads. No manifest edit alone resolves these.

ok already accounts for --strict: it is false when there are hard errors, and also when --strict is passed and there are warnings. The process exit code matches, and the JSON goes to stdout while the failure is reported on stderr β€” so civitai app validate --json | jq works on a project that fails validation.

πŸ”΄ BREAKING β€” a refused path now emits no object at all. This object is written only when validation actually produced a result. A path that does not exist, or that is not a directory, is a mistake about the invocation: it writes nothing to stdout and exits 2. It used to print {"ok": false, "dir": "/nope", "errors": [ … ]} and exit 1 β€” a fabricated validation result, complete with a finding about a manifest nobody could have written. A failure that produces no validation result at all likewise emits no object and exits 1 β€” a project directory the CLI cannot stat (it is unreadable, or a component of the path below it is not a directory), and in principle an internal schema failure, which is a directory the CLI can read that still yields nothing to print.

An unreadable block.manifest.json is not one of those cases β€” it is a validation verdict, and the object is printed in full with a single (root) finding carrying the permission denied message. The distinction is how far the CLI got before it stopped: it could not read your manifest, which is something to report about the project; it could not read the directory, which is nothing at all.

So branch on the exit code before parsing:

exit stdout
0 the object, "ok": true
1 the object with "ok": false for a validation verdict β€” including an unreadable manifest β€” but nothing when validation produced no result at all (an unreadable project directory, or an ENOTDIR partway down the path; also an internal schema failure, which a released binary should never hit)
2 nothing β€” the path does not exist, or is not a directory

πŸ”΄ jq -e is the wrong tool for reading ok. It exits 1 on a JSON false, which is indistinguishable from its exit code for a missing key β€” so the obvious one-liner reports a failing but perfectly well-formed result as "no result", which is the one distinction this whole section exists to draw. Test for an empty string instead, and read ok as a value:

out=$(civitai app validate ./my-block --json); rc=$?
case $rc in
  2) echo "bad path β€” check the argument"; exit 2 ;;
  0|1)
    if [ -z "$out" ]; then
      echo "no result to parse (rc=$rc)"; exit "$rc"
    fi
    echo "ok=$(jq -r .ok <<<"$out")"          # true | false β€” the verdict
    jq -r '.errors[]   | "ERROR   \(.field): \(.message)"' <<<"$out"
    jq -r '.warnings[] | "WARNING \(.field): \(.message)"' <<<"$out"
    ;;
esac

Submit & auth

civitai login (no flags) runs the OAuth device-authorization grant: it prints a URL + a short code, you approve in your browser, and the CLI stores a short-lived access token (1h) plus a refresh token (30d) that it rotates automatically before requests and once on a 401. By default it requests UserRead | AppBlocksSubmit | AppBlocksDevTunnel (== 100663297) β€” identity, Apps submit (which gates both app submit and the dev-token mint), and the on-site dev tunnel. That default deliberately omits AIServicesWrite, so a plain login cannot spend Buzz: it drives the read/identity dev:live paths (viewer, catalog, app storage) but, for a generation app, has its ai:write:budgeted scope stripped at mint time and cannot estimate, submit, or spend real Buzz.

Opt into generation explicitly with a named scope set:

civitai login --scopes generate   # additive: keeps submit + dev-tunnel, ADDS generation

which requests 100777985 (the default plus AIServicesRead | AIServicesWrite | BuzzRead) β€” one credential that can both submit apps and run civitai generate. --scopes takes a named set, never a raw bitmask, and an unknown name is rejected with the valid list. It applies only to the browser device login and is refused alongside --token.

πŸ”΄ The device-flow scope check is all-or-nothing: requesting any bit the civitai-cli OAuth client's allowedScopes does not permit rejects the whole login with invalid_scope. On civitai.com that client's allowedScopes is 100777985, so --scopes generate works. Against a self-hosted or older auth server (a non-default CIVITAI_BASE_URL) that predates the widening it is rejected, and the CLI maps the rejection to a message telling you plain civitai login still works there.

A full-scope personal API key remains the other way to get spend authority (see the credential table under Local dev loop above), and is still the only credential carrying the rest of the Full scope mask. civitai login --token <key> stores a personal API key instead (no refresh). CIVITAI_TOKEN overrides the stored credential (treated as a personal key).

civitai app submit:

  • always validates + packages the canonical source ZIP, then
  • uploads it with your stored token to the token-authenticated route POST /api/v1/blocks/submit-version (Authorization: Bearer). OAuth tokens refresh transparently. Set CIVITAI_SUBMIT_PATH to override the route.
  • With no token configured (and not --package-only), it instead writes the .zip and prints the next steps (civitai login, or web upload at /apps/submit).

--package-only always just writes the .zip and stops.

After you submit: review β†’ approve β†’ deploy

A successful submit does not publish your app β€” it queues it for moderator review. The lifecycle is:

  1. submit β†’ your submission lands at /apps/my-submissions with status pending.
  2. review β†’ a moderator reviews the manifest + files. They either approve or reject (with a reason you can read inline, then fix and resubmit).
  3. deploy β†’ on approval, the platform builds and deploys your app (injects its build recipe β†’ builds the image β†’ deploys β†’ programs the <blockId>.civit.ai DNS record). A few minutes after approval it serves live at https://<blockId>.civit.ai/.

Before approval, https://<blockId>.civit.ai/ 404s β€” submitting does not make the subdomain serve (but dev:live works against a pending app β€” see Local dev loop). For the full end-to-end walkthrough (build β†’ submit β†’ review β†’ deploy), see the Build your first App guide.

Budget time for artwork. Review and deploy are not the only gate: a store listing cannot go live without an icon AND a cover. civitai app submit mints your listing as a draft and prints this reminder inline β€” which is the moment to act on it, because the media is settable while the app is in review and carries forward when a moderator approves it. Attach it without the browser:

$ civitai app listing status                        # what's attached vs. required
App:             my-app
Listing status:  draft
Icon:            MISSING (required)
Cover:           MISSING (required)
Screenshots:     0

⚠ Not publishable yet β€” missing icon and cover.
  Add one:  civitai app listing set-icon <file>
  Add one:  civitai app listing set-cover <file>

$ civitai app listing set-icon ./assets/icon.png    # png/jpeg/webp, ≀2 MiB
$ civitai app listing set-cover ./assets/cover.png  # png/jpeg/webp, ≀4 MiB
$ civitai app listing add-screenshot ./assets/screenshot-1.png --caption "Grid view"   # optional, up to 8

assets/ is scaffolded by every template, with a README of the requirements and no placeholder images β€” the files above are ones you supply.

Details worth knowing before you start:

  • Source images are png/jpeg/webp and are size-checked locally before any upload β€” icon ≀2 MiB, cover ≀4 MiB, screenshot ≀2 MiB. That is the whole of what the CLI enforces.
  • Dimension and aspect rules are the platform's, and it states them. The CLI does not enforce them locally β€” a copied number goes stale and starts refusing valid images β€” but it does document the current bounds, as guidance rather than as a gate: see Listing media requirements. It uploads, attaches, and then waits for the content scan β€” in that order, because the platform validates dimensions, aspect and format at the attach step. So a wrongly-shaped image comes back in a couple of seconds with the platform's own message naming the bound and your value (e.g. icon must be square-ish (aspect 2.00 outside 0.9–1.1)), instead of after the scan has finished.
  • A blocked image never goes live. The scan verdict is still waited on, so these commands never report success on a pending or blocked scan; a failure tells you what state the listing was left in.
  • The app is resolved from block.manifest.json in the current directory; pass --slug <blockId> (with --dir if you prefer) to run it from anywhere.
  • On a listing that is already LIVE, attaching media does not edit the live listing β€” it opens a revision that goes back to moderator review. Describe it with --changelog "<what changed>"; -y/--yes skips the confirmation. civitai app listing status on a live listing reports that in-progress revision's media, and tells you when a revision is already under review.
  • Screenshots are managed by id (the alsc_… ids listing status prints): rm-screenshot <id> removes one, and reorder <id...> takes all the current ids in the new order β€” a partial set is rejected.

Need to change the bundle while a request is still pending? Withdraw it first to free the slug, then resubmit:

$ civitai app status                          # find the pubreq_ id
$ civitai app withdraw pubreq_01HZX           # frees the slug
$ civitai app submit                          # resubmit the new bundle

civitai app withdraw <pubreq-id> (or --id <pubreq>) withdraws your own pending publish request. It is idempotent (an already-withdrawn request still returns success) and only a pending request can be withdrawn β€” an already approved/rejected one cannot.

Listing media requirements

Two different things check your listing images, and it is worth knowing which is which before you open an image editor.

What the CLI checks, locally, before anything is uploaded: the file's format and its byte size. That is all.

kind how many format byte cap (the file you pass)
icon 1, required png / jpeg / webp ≀ 2 MiB
cover 1, required png / jpeg / webp ≀ 4 MiB
screenshot up to 8, optional png / jpeg / webp ≀ 2 MiB

What the platform checks, server-side, when the image is attached: the dimensions and the aspect ratio. The CLI does not reproduce these, so this table is guidance β€” the server is the authority and its rejection names the bound and your value (icon must be square-ish (aspect 2.00 outside 0.9–1.1)).

kind aspect (width Γ· height) minimum size
icon 0.9 – 1.1 β€” square-ish, not exactly square 128 px on the shorter side
cover 1.3 – 2.4 β€” landscape, ~4:3 to ~21:9 640 px wide
screenshot 0.4 – 2.6 β€” either orientation 320 px on the shorter side

Easy starting points: a 512 Γ— 512 icon and a 1600 Γ— 900 cover.

Four behaviours that are not obvious from the numbers:

  • Icons are re-encoded server-side. Whatever you upload is downscaled to at most 1024 px on its longer side and re-encoded to PNG β€” aspect preserved, and never enlarged. So an oversized icon is harmless, but an undersized one is not: the 128 px floor still bites, because nothing is ever scaled up. One consequence to know about: the platform also caps the re-encoded icon at 1 MiB, and that is a different measurement from the 2 MiB the CLI applies to the file you pass. A detailed photographic icon can clear the local check and still be refused for the size of the PNG the server made from it β€” the message quotes bytes, not pixels. Flat, simple artwork re-encodes far smaller.
  • An icon's upper bound is a PIXEL count, not a file size. The decoder that re-encodes it refuses a source above roughly 16 megapixels β€” about 4096 Γ— 4096 β€” and it refuses it regardless of how small the file is. A flat 5000 Γ— 5000 PNG compresses to a few hundred KB, so it clears every byte cap in the first table and is still rejected. Downscale before you upload: 1024 Γ— 1024 is plenty, because that is what the server re-encodes to anyway.
  • Covers and screenshots are not rescaled. What you upload is what the store renders, so ship them at the size you want shown.
  • A wrong image is rejected, not quietly accepted, and it comes back fast. The CLI attaches before it waits on the content scan, so the platform's verdict on shape arrives in a couple of seconds rather than after a scan that can take two minutes β€” and always before a moderator sees it. The message names the bound it applied and the value it measured β€” read it rather than guessing which limit you crossed.

Why the CLI does not enforce the second table. These are platform constants that can move. Stale guidance costs you one rejection that carries the current bound; a stale local gate refuses valid images and cannot be argued with. The asymmetry is the reason the split exists β€” please don't "helpfully" promote these numbers into a local check (see AGENTS.md item 25).

Submission status

civitai app status checks where your own submissions are in that lifecycle without leaving the terminal. It calls the token-authenticated, self-scoped route GET /api/v1/blocks/submissions with your stored credential β€” you only ever see your own submissions (the same token that submitted can read its status; OAuth tokens need the Apps submit scope).

With no argument it lists every submission, newest first:

$ civitai app status
BLOCK_ID    VERSION  STATUS    DEPLOY    SUBMITTED   URL
gen-matrix  0.6.0    approved  live      2026-06-22  https://gen-matrix.civit.ai/
my-block    0.2.0    pending   -         2026-06-21  -
old-app     0.1.0    approved  building  2026-06-19  -

Pass a blockId (app slug) or --id <pubreq_id> to see one in detail β€” including the rejection reason if it was rejected (so you can fix + resubmit) and the live URL once it is approved and deployed:

$ civitai app status gen-matrix
Block ID:         gen-matrix
Version:          0.6.0
Publish request:  pubreq_01HZX
Status:           rejected
Deploy state:     -
Submitted:        2026-06-22 09:05 CDT
Reviewed:         2026-06-22 11:40 CDT

Rejection reason:
  the budgeted scope needs the per-app Sybil cap signed off first

Not live yet β€” gen-matrix.civit.ai only serves after the app is approved and deployed (deployState 'live').

The unfiltered listing is capped server-side at 100 rows, and the API returns no cursor and no total β€” so there is no way to page and no way to know how many were dropped. When a full-length page comes back the CLI says so on stderr rather than presenting it as your complete history:

note: showing the newest 100 submissions β€” the API caps this listing and offers no way to page, so older submissions may exist but are not listed. Look up a specific app with `civitai app status <blockId>`.

That is an inference (a page that is exactly full is indistinguishable from one that was cut off), so it says may. A per-app lookup β€” civitai app status <blockId> β€” is not affected: the server narrows to the slug before applying the cap.

--json emits the raw response for scripting. An empty list prints a friendly "run civitai app submit" hint; with no token it points you at civitai login. Notes like the cap caveat go to stderr, so --json stdout stays pure and the exit code stays 0.

Deployed is not the same as listed in the store

civitai app status and civitai app view read different resources, and an app can legitimately be in one and not the other:

  • civitai app status <slug> reads your submission pipeline (GET /api/v1/blocks/submissions) β€” review status, deploy state, live URL.
  • civitai app view <slug> reads the public store catalog (GET /api/v1/apps/{slug}) β€” the published store listing.

So app status can show approved / live with a working <slug>.civit.ai URL while app view <slug> returns not found (exit 4). That 404 is truthful and says nothing about your deploy: the store lists an app only once its store listing is published (a listing needs an icon and a cover β€” see civitai app listing status), and the catalog itself is still gated by a launch flag while the store is pre-GA. When the 404 lands on a slug you own, the CLI detects that and says so, naming both next commands, instead of leaving you with a bare "App not found".

Pull your app's repository (app pull)

civitai app pull is the read side of git authoring: it clones (or, if [dir] is already a checkout, syncs) the canonical repository backing one of your approved Apps, so you can edit locally and then civitai app submit or push.

civitai app pull --app my-block                # clone into ./my-block
civitai app pull ./my-block --app my-block     # clone/sync into ./my-block
civitai app pull . --app my-block              # sync the current directory

--app is required (the slug or appBlockId; find it with civitai app status). Authentication uses your stored credential (civitai login or a personal API key) β€” the command calls an owner-only endpoint that lazily provisions a scoped, read-only Forgejo identity for you and returns a clone URL with a pull token embedded.

The repo only exists once your first version has been submitted as a ZIP and approved; before then the command tells you so rather than failing obscurely.

⚠️ SECURITY β€” TOKEN-IN-URL LEAKAGE. The clone URL embeds your access token as HTTP-Basic credentials (https://<user>:<token>@…).

  • On a fresh CLONE, git writes the remote URL into .git/config, so the token lands on disk in the checkout. Treat the directory as sensitive: do not commit .git/config or share the directory. To drop the token, point the remote at the credential-less HTTPS URL β€” git -C <dir> remote set-url origin <httpUrl> β€” which is exactly the command the CLI prints for you after a clone.
  • On a SYNC (pull into an existing checkout) the URL is passed explicitly and is not persisted to .git/config. The token still appears transiently in the git child process's arguments, so it is briefly visible to other local processes via ps / /proc/<pid>/cmdline.

The CLI prints the applicable half of this warning to stderr after every run.

A sync is git fetch + git merge --ff-only, so it never creates a merge commit and refuses rather than clobbering diverged history or a conflicting dirty tree.

Browse the App store

civitai app list and civitai app view <slug> read the public App store catalog (GET /api/v1/apps and GET /api/v1/apps/{slug}).

πŸ”΄ These are not anonymous reads. Unlike the model/image/article commands in Browse the public API, both require a credential and exit 3 with no token configured β€” run 'civitai login' (or set CIVITAI_TOKEN) to browse the App store without one. The store endpoint keys the visible catalog off your identity, so an anonymous call would see nothing; the CLI refuses up front rather than presenting an empty catalog as the whole store.

civitai app list
civitai app list --kind onsite --sort popular --limit 10
civitai app list --category generation --json
civitai app list --cursor '<next-cursor-from-a-previous-page>'
civitai app view my-cool-app
civitai app view my-cool-app --json
  • Filter-based discovery, not search. There is no free-text query β€” the store service does not expose one, which is why there is no civitai app search. Filter with --kind (all, onsite, offsite), --category (generation, games, utility, discovery, moderation, analytics, other) and --sort (top-rated, popular, newest, name).
  • Keyset cursor pagination, not page numbers: the next cursor is printed after the results, and you pass it back with --cursor. --limit is 1–50.
  • The store is gated by a launch flag. Until it opens publicly you only see apps if your account is a moderator or an app-dev-tester, so a perfectly valid login may get an empty list. That is the pre-GA state, not a broken login.
  • Rate-limited per caller β€” a tight scripted loop may see 429s, which the CLI backs off and retries automatically.
  • app view reads the store catalog, which is not your deploy: see Deployed is not the same as listed in the store.

App metrics

civitai app metrics <slug> shows the owner-only analytics for one of your App Blocks. The slug is resolved to its appBlockId through your own submissions, so analytics exist only once a version has been approved β€” an app still in review reports that instead of an empty dashboard.

$ civitai app metrics gen-matrix --from 2026-05-01 --to 2026-08-03
App:          gen-matrix
Window:       2026-05-01 00:00 UTC β†’ 2026-08-03 00:00 UTC
Granularity:  week

Installs
  Total   12
  Active  9

Runs
  Count       20
  Buzz spent  65

Buzz purchased
  Purchases  3
  Buzz       15000
  Gross      $14.97

App loads
  Impressions       124
  Unique viewers    12
  Signed-out loads  40

Engagement
  API calls     26
  Active users  2
  Error rate    3.8%

  Top scopes:
    ai:write:budgeted  20

  Top endpoints:
    /api/v1/blocks/me  4

Three things are worth knowing, because each one otherwise produces a believable-but-wrong reading:

  • The window is always printed, and it comes from the server. The API defaults to the last 30 days and clamps any request to 366 days, so a real app with 20 runs in mid-June reads 0 under the default window. The window shown is the one the server actually served β€” if it clamped your --from, the printed range says so. Widen it with --from / --to, which accept a plain YYYY-MM-DD (midnight UTC) or a full RFC3339 timestamp; a malformed value or an inverted window is a usage error (exit 2) caught before any request.
  • "Not entitled" is not "zero". When the caller doesn't own the app (or lacks Apps-author access) the API answers HTTP 200 with every counter zeroed, flagged only by a notOwned field. The CLI refuses to render a dashboard in that case and tells you to check civitai whoami / civitai app status <slug> instead β€” a silently-empty dashboard that looks like real data is the failure mode this command is built to avoid.
  • It needs a personal API key. The query is full-scope, so an OAuth civitai login token gets a 403. Both refusals β€” the 403 and the no-token-at-all case β€” name the route that actually works (civitai login --token <key>, created at civitai.com/user/account) rather than the generic "run civitai login", which here is the one route that cannot succeed.

Two data caveats.

Engagement counts only authenticated, scope-gated API calls. An app that ships no scoped API surface shows real installs and revenue with a flat engagement section β€” that is expected, not a bug. Installs is a different case again: it shows n/a for an app that cannot be installed at all (a page app has no install slot, so an install record cannot exist), which is deliberately distinct from a real 0 on an installable app nobody has installed yet. App loads is the exception: it is measured on every load, so it counts signed-out visitors and static blocks that engagement structurally cannot see. Unique viewers counts signed-in people once each and approximates signed-out ones by network address, so read it as reach rather than an identity count, and Signed-out loads is a count of LOADS (one anonymous visitor reloading ten times is 10 there and 1 unique viewer), so it can legitimately exceed Unique viewers. Note also that these are mount ATTEMPTS: a load that FAILED still counts, because a failed mount's only beacon is the same event and it carries no status. Error rate is the share of those calls that failed, and the human view renders it as a percentage (the server sends it as a 0–1 ratio, which --json passes through unchanged). Only a genuine zero prints 0.0%: a real but tiny rate β€” a high-traffic app with a handful of failures β€” reads <0.1% rather than rounding away to look error-free.

--json emits the raw analytics payload (the server's own object, including notOwned and the per-bucket series arrays the human view omits) for scripting. Note that --json does not refuse a not-entitled read the way the human view does: a notOwned: true payload is passed through with every counter zeroed and the command still exits 0, so civitai app metrics <slug> --json | jq .runs.count returns 0 for an app you can't see. A script must branch on the notOwned field rather than trusting the counts.

Installs carries installs.notApplicable for the case above. It is NOT an outage flag β€” it means the question does not apply to this app type, so a script should render it as "not applicable" rather than retrying or warning about infrastructure. --json passes it through and still exits 0, so branch on it rather than trusting the counts.

App loads has a SECOND, section-local unavailability flag β€” views.unavailable, independent of notOwned. It is the one section the server reads from a different store, which can be unreadable β€” or merely too slow, the read is time-bounded server-side β€” while every other counter in the same response is genuinely measured. When that happens the human view prints unavailable and says so explicitly rather than printing a 0 you would read as "nobody opened my app". --json passes the flag through and still exits 0, so a script must branch on views.unavailable too β€” jq .views.count alone cannot tell an outage from a real zero. A server old enough to predate this section omits the views key entirely; the human view reports that as unavailable as well (naming the different cause), and a script should treat a missing .views the same way.

Generate

civitai generate "<prompt>" runs a text-to-image generation on Civitai's generator.

πŸ”΄ This spends real Buzz and cannot be undone. A submitted generation is charged the moment the orchestrator accepts it, and nothing local calls that back β€” not --timeout, not Ctrl-C, not civitai workflows cancel. Price it with --dry-run first β€” that calls the cost estimator and spends nothing.

πŸ”΄ What the LEDGER does with a charge is not something this CLI reports, in either direction. If a run fails, expires, or you cancel it, whether any Buzz comes back is decided server-side; this CLI cannot see your Buzz ledger β€” civitai buzz reports a balance, not a history, so settle it against your Buzz transaction history (/user/transactions). The CLI states the outcome and stops there. It does not tell you the charge stands, and it does not tell you it was refunded β€” earlier versions asserted the first, which the platform's own client contradicts for failed/expired/canceled. The rule itself lives in the orchestrator service and is not readable from the civitai monorepo, so neither claim is made. Tracked at civitai/cli#307.

# Price it. Spends nothing.
civitai generate "a cat wearing sunglasses" --dry-run

# The same estimate as raw JSON, for scripts
civitai generate "a cat wearing sunglasses" --dry-run --json

# Generate, refusing if the estimate exceeds 50 Buzz
civitai generate "a cat wearing sunglasses" --quantity 4 --max-cost 50

# A specific checkpoint plus a LoRA at 0.8 strength
civitai generate "a cat" --checkpoint 128713 --lora 250712:0.8

# Wait for the result and write the images into ./out
civitai generate "a cat" --yes --out-dir ./out

# Fire and forget; collect the results later
civitai generate "a cat" --yes --no-wait
civitai workflows list
civitai workflows get <workflow-id>

# Non-interactive (CI) β€” --yes is required, or the run is refused
civitai generate "a cat" --yes --max-cost 20

# Image-to-image from a local file β€” --ecosystem is REQUIRED with --image
civitai generate "make it winter" --ecosystem Flux1Kontext --image ./cat.png --dry-run

# …or from a public https URL, with two reference images
civitai generate "combine these" --ecosystem Seedream \
  --image https://example.com/a.jpg --image ./b.png --yes

# Graduate from flags to a raw graph: print, edit, send back
civitai generate "a cat" --quantity 2 --print-input > graph.json
civitai generate --input graph.json --dry-run

Credential. Generation needs the AI Services scopes. Either a full-scope personal API key (create one, then civitai login --token <key>), or a browser login that opted in: civitai login --scopes generate. A default OAuth browser login (civitai login, no --scopes) does not carry them and is refused. civitai whoami shows the capability as Spend Buzz (AI Services).

The one exception is --print-input: it assembles the graph and exits before the estimator, the submit and the balance read, so it needs no credential β€” useful for building a graph to edit before you have logged in. Two caveats, and they are not the same caveat:

  • --print-input with --image does need a credential, because it uploads each local file first and that upload is authenticated.
  • --print-input with --checkpoint or --lora needs none β€” the model-version lookup is a public read β€” but it is not offline: that lookup is a real request, and with no network it fails (exit 5) rather than printing a graph. Measured against a dead endpoint: bare --print-input exits 0; --print-input --checkpoint <id> exits 5 after the read's retries.

So only a bare --print-input needs neither a credential nor a network.

πŸ”΄ --max-cost is an estimate check, not a spending cap

The cost this command shows is an estimate, not a quote: the server's estimator returns no quote id, no signed price and no expiry, so there is nothing to hand back at submit time β€” and no server-side spending ceiling is reachable from an API key at all. The realized charge can exceed the estimate, and --max-cost cannot claw the difference back β€” it never reaches the server. What the ledger then does with that charge is not something this CLI reports β€” see What the LEDGER does with a charge under Generate.

--max-cost compares that estimate against your number and refuses locally before submitting. It catches a --quantity typo. That is all it can do. Do not run an unattended loop believing it caps spend. (The per-API-key buzzLimit on your account does not bind this path either β€” the generator meters a separate server-minted subject, not your key.)

πŸ”΄ Silent model substitution

If you pass a --checkpoint version id that is not valid for the model family being generated, the server does not reject it. It substitutes that family's default checkpoint, runs the job, and bills you for what actually ran. Until recently the reply was indistinguishable from success β€” a nonexistent version id came back 200 OK at the default price.

The server now reports each swap, and civitai generate surfaces it:

$ civitai generate "a cat" --checkpoint 999999999 --dry-run
⚠ The server will NOT use the checkpoint you asked for. It has substituted a
  different model, and this estimate prices the SUBSTITUTE. Nothing has
  been submitted or charged yet.
    requested version 999999999 -> will run version 2436219  (reason: unrecognized)
      the server does not offer that version in this model family at all …

The checkpoint line in the summary you approve is annotated too, so the model that will not run is never the last one you read before saying yes:

Checkpoint:   DreamShaper β€” 8 (Checkpoint, id 128713)  [SUPERSEDED β€” the server will run version 2436219 instead; see the warning above]

Your own id stays on the line: the CLI marks it, it never quietly substitutes the applied one.

It is reported on the estimate (--dry-run, and before the confirmation prompt on a real run β€” while you can still back out), again after the submit (where it is the receipt for what was billed), and on a later read with civitai workflows get <id>, which is the only place a --no-wait run can still discover it. The report always goes to stderr, so --json stdout stays machine-clean β€” and --json carries the raw modelSubstitutions array itself.

There are three reason values, and they want different fixes:

reason What it means What to do
wrong-workflow The version is real for this family but scoped to a different workflow (e.g. an edit-only version sent to text-to-image). Pick a version offered for the workflow you are running, or change --ecosystem to match.
unrecognized The version is in no list for this family β€” a community checkpoint, or one retired since your script was written. Check it with civitai model-versions get <id> and pin one that is still offered.
gated The version is offered here, but a gate rule hides it from your account. An entitlement issue, not a command mistake: may need a membership, an early-access window, or an accepted licence.

By default this is a warning and the run continues β€” the substitution is a deliberate graceful degradation, so a script pinned to a version that was later retired keeps working rather than breaking on a CLI upgrade. Pass --fail-on-substitution to refuse instead; it is checked against the estimate, so nothing is submitted and nothing is charged when it refuses.

πŸ”΄ Silence is not an assurance. The field is omitted when nothing was substituted, so "no warning" means either "no substitution" or "a server older than this feature" β€” the CLI cannot tell those apart and deliberately never claims the negative.

πŸ”΄ --fail-on-substitution is therefore NOT a spend guard. It can only refuse what the server reports. Against a deployment that does not report substitutions the flag is silently inert β€” exit 0, submitted, charged β€” and there is no signal distinguishing that from "nothing was substituted". Two further limits worth knowing before you build a pipeline on it:

  • It is evaluated on the estimate. A substitution appearing only on the submit reply is reported (and is in --json) but exits 0, because the money is already gone and failing there would strand a result you paid for and still need to collect.
  • It refuses on the first reported substitution; the message names that one and counts the rest.
Confirmation

An interactive run prints the estimate, your balance and the resolved model names, then asks. A non-interactive shell (pipe/CI) without --yes is refused rather than charged silently. Everything the confirmation prints goes to stderr, so --json keeps stdout machine-clean.

Image-to-image: --image and --ecosystem

--image <path-or-url> (repeatable) attaches a reference image and turns the job into an edit. A local .png/.jpg is uploaded to Civitai first and the stored blob is referenced; an https URL is passed through as-is, but must be publicly reachable β€” the generator downloads it server-side too, and an unfetchable URL is a 400 after you have already been priced. Either way the CLI reads the image's width and height from its header only (never decoding the pixels) and sends them, because the server requires both and rejects an entry without them. http://, file:// and data: are refused, local files are capped at 64 MiB (checked by stat, before a byte is read), and only png and jpeg are supported β€” webp would need a new third-party decoder dependency.

πŸ”΄ --image requires --ecosystem, and the reason is money. The server promotes a text-to-image job to image-to-image only when the request names an ecosystem. Without one it ignores the images, generates from the prompt alone, and charges you the full amount β€” HTTP 200, no error, no warning. Measured: the same graph with and without images[] priced byte-identically.

Two more things the CLI genuinely cannot check for you, so it says them instead of pretending:

  • Only some ecosystems accept reference images at all. Qwen, Flux1Kontext, NanoBanana, Seedream, OpenAI, Grok, Reve, MAI, Boogu and a few more do; the Stable Diffusion family and the default ecosystem do not β€” and for those the images are dropped silently and billed. The cost estimate cannot tell you which case you are in: several edit-capable ecosystems price identically with and without images (measured on Flux1Kontext, NanoBanana and Seedream), so a price comparison is not a detector. Name an ecosystem you know supports editing.
  • Too many reference images are silently truncated. Per-ecosystem limits run from 1 to 7, live only inside the server's per-engine graphs, and the extras are dropped before any limit check can fire β€” so the server never reports it and the truncated job is billed. Measured on Qwen (limit 3): 4, 5, 6 and 12 images all priced identically to 3. The CLI refuses more than 7 (no ecosystem accepts more, so that refusal can never block a valid request) and warns for anything above 1. It deliberately does not vendor the per-ecosystem table β€” see AGENTS.md items 13 and 19(c).

--ecosystem is sent to the server verbatim and is not checked locally; an unknown value comes back as the server's own unknown ecosystem error.

--dry-run does upload local --image files, because an estimate built on a graph with no images[] prices a plain text-to-image job. Uploading spends no Buzz, and --dry-run still never submits.

The content flags, and why there aren't twelve

--negative-prompt, --quantity, --aspect-ratio, --checkpoint <version-id>, --lora <version-id>[:strength] (repeatable), plus --image / --ecosystem above.

The generator is permissive, not a validator β€” it returns HTTP 200 for things it silently changes:

  • An out-of-range --quantity is clamped with no error (asking for 40 charges you for the server's limit). The CLI warns when you cross it.
  • --steps 0 / --cfg-scale 0 are accepted and price a degenerate, cheaper, wrong job β€” which is exactly why those flags are not exposed yet.
  • A checkpoint id that does not exist is accepted, the ecosystem default is silently substituted, and you are billed for it.

So --checkpoint and every --lora is resolved against the public model-version API before anything is submitted: a bad id becomes a hard local not found (exit 4) instead of a wrong charge, and the confirmation echoes the resolved model name so you approve a name rather than an integer. --model is deliberately absent: civitai download --model takes a model id, while this takes a version id.

Raw graphs: --print-input and --input

The five flags cover the common job. Everything else the generator understands lives in the generation graph β€” the JSON document the flags assemble. You can write that document yourself:

# 1. Assemble it from flags, print it, and exit. No submit, no cost estimate,
#    no balance read β€” with no --checkpoint/--lora, no request at all.
civitai generate "a cat" --quantity 2 --aspect-ratio 1:1 --print-input > graph.json

# 2. Edit graph.json however you like.

# 3. Send it as-is. Price it first; --dry-run still spends nothing.
civitai generate --input graph.json --dry-run
civitai generate --input graph.json --yes

# …or pipe it, with `-`
jq '.prompt = "a dog"' graph.json | civitai generate --input - --dry-run

--print-input reaches no money seam: not the submit, not the cost estimator, not the balance read. With --checkpoint/--lora it does still make the public model-version read those flags always make β€” that lookup supplies model.type, which graph resources[] require, so skipping it would print a document --input could not submit.

--print-input's output is a valid --input document by construction β€” that round-trip is the point of the pair, and it is what replaces a --set some.path=value expression language the CLI deliberately does not have (a wrong type in such an expression is accepted by the server silently, and billed; an edited file is inspectable before it is sent).

Four things to know, all of them consequences of it being a passthrough:

  • txt2img only. A graph declaring any other workflow is refused. The server's content audit reads the top-level prompt node, and it rebuilds what it inspects from declared graph nodes β€” so a graph carrying its prompt somewhere else (a comfy node, a nested step input) is exactly the shape that could reach the generator unaudited. That question is open upstream, and this CLI will not be the path that answers it the wrong way.
  • Envelope keys are refused, not ignored. civitaiTip, creatorTip, buzzType, tags, externalId, sourceMetadata, sourceMetadataMap, remixOfId and a top-level input belong to the request envelope around the graph, not to the graph. A file setting civitaiTip would charge a tip that --dry-run structurally cannot show you β€” the estimator prices a strictly smaller request and is never sent tips at all β€” so the file is rejected with an error rather than quietly cleaned up.
  • Keys the CLI does not model are passed through, with a warning. The warning says the CLI cannot verify the key; it is not a claim that the key is invalid, because the CLI does not carry a copy of the server's node registry. It matters because the server's failure mode for a key it does not declare is to drop it silently at HTTP 200 β€” a typo costs Buzz and produces a job that ran without your parameter, with no error anywhere.
  • No model-id safety net. --checkpoint / --lora are resolved against the public API before submitting; a raw graph is not interpreted, so a nonexistent id in it is accepted, the ecosystem default is substituted, and you are billed.

--input cannot be combined with a prompt argument or with --negative-prompt / --quantity / --aspect-ratio / --checkpoint / --lora β€” there is no predictable answer to "does --lora append to or replace the file's resources?", so the combination is a usage error. Every execution flag (--dry-run, --yes, --max-cost, --json, --no-wait, --timeout, --out-dir, --no-download, --force, --external-id) still applies.

Waiting, downloading, and re-attaching

By default generate waits for the job to finish and writes every deliverable output into --out-dir (default .) as <workflow-id>-<n>.<ext>. --force overwrites existing files; without it a collision is refused before any bytes move.

  • --no-wait submits, prints the workflow id and exits 0.
  • --no-download waits and prints the output URLs instead of writing files.
  • civitai workflows get <workflow-id> shows a workflow at any time. It is the re-attach path for every case where the CLI stopped early, and it spends nothing.

πŸ”΄ --timeout stops waiting. It does not stop paying. When the deadline passes (or you press Ctrl-C) the generation keeps running server-side, and cancelling it does not stop the cost already accrued β€” a mid-run cancel bills that. Both cases exit non-zero, print the workflow id, the idempotency key and the exact civitai workflows get … command, and never report success.

Output URLs are presigned and expire. Download promptly; re-read the workflow for fresh links. The blob fetch deliberately carries no credential β€” the URL is already authorized, and attaching your full-scope API key to it would hand 25 unrelated permissions to a request that needs none.

A finished workflow can contain fewer usable results than you paid for. An output can be blocked by moderation, never land, or be one you hid on the website. Those are filtered out of the download β€” and reported, with the reason, plus an explicit note when the count differs from --quantity. Silently writing three files for a four-image job is the failure this exists to prevent. If every output is filtered out the command exits non-zero.

Crash safety. The orchestrator's idempotency key is written to ~/.config/civitai/pending/<key>.json before the request is sent, because the money moves server-side even if the process dies mid-POST. If a submit's reply never arrives, re-run with --external-id <key>: the orchestrator dedupes on it and returns the pre-existing workflow instead of charging again (it answers a duplicate with HTTP 200, not a 409, so re-attachment is inferred locally).

Polling cadence. The status poll starts at 5s, backs off exponentially to a cap, and backs off harder on a 429. That floor is not tunable downward: the workflow read proxies straight through to the orchestrator with no cache and no server-side rate limit, so the CLI's own restraint is the only thing between it and a 429 storm.

Listing and cancelling workflows
civitai workflows list                      # newest first
civitai workflows list --limit 5
civitai workflows list --limit 50 --cursor <next-cursor>
civitai workflows list --json               # raw server payload, incl. nextCursor
civitai workflows cancel <workflow-id>      # asks for confirmation
civitai workflows cancel <workflow-id> -y   # skip the prompt (scripts/CI)

list is cursor-paged, not page-numbered: when more results exist it prints Next cursor: <c> on stdout, which you pass back as --cursor. --tag filters on orchestrator workflow tags (repeatable).

The OUTPUTS column reads deliverable/total. The two differ when an output was blocked by moderation, never landed, or you hid it on the website β€” so 0/4 means four images were produced and paid for and none of them are usable, which is a very different fact from 0/0. civitai workflows get <id> shows the per-output reason.

πŸ”΄ cancel does not undo the charge. A mid-run cancel bills the cost already accrued, orchestrator-side: by the time a workflow is running the money has moved, and stopping it does not call that back. Cancel a job because you no longer want its output β€” never as a way to save Buzz, and never to undo a submit. Whether the ledger returns anything afterwards is the server's call and this CLI does not report it either way (civitai/cli#307). (This is also why --timeout and Ctrl-C deliberately do not cancel: stopping the wait costs nothing, while stopping the job would cost the same as letting it finish and throw the result away.)

cancel asks for confirmation, matching civitai generate and civitai app submit. It is the one irreversible action here, and it destroys a job you have already paid for, so it is gated the same way every other destructive path in this feature is:

  • --yes / -y proceeds without prompting;
  • an interactive terminal prints what is lost and prompts β€” the default is no, so a bare Enter aborts;
  • a non-interactive shell without --yes refuses rather than cancelling silently. Scripts must pass --yes explicitly.

Nothing is cancelled when the confirmation is refused β€” the gate runs before the request goes out.

Exit codes specific to generate

generate follows the global exit-code table, with one deliberate refinement. The API answers several very different failures with the same HTTP status, and the generic mapping would send a script down the wrong path β€” in particular a caller who is out of Buzz, muted, or hitting a server-side outage must never be told to re-run civitai login. Those cases therefore exit 1 (generic), not 3 (auth) or 2 (usage):

πŸ”΄ An exit code does not tell you whether you were charged. Every failure above the divider happens before anything is submitted, so nothing was spent. Every failure below it happens after the submit, and the Buzz is gone β€” including a --timeout, a Ctrl-C, and a workflow that ends failed. Do not write a retry loop that branches on the exit code alone; re-attach with civitai workflows get <workflow-id> instead of re-submitting.

Failure Exit
β€” nothing submitted, nothing spent β€”
Missing AI Services scope / no token / not authenticated 3
Not enough Buzz (caught locally against your balance, or reported by the server) 1
Account muted, or onboarding incomplete 1
Generation disabled server-side 1
Prompt refused by content moderation β€” πŸ”΄ never retry, repeated blocked prompts get the account muted 1
The server priced the job but reports ready: false (a selected resource is not currently available) 2
Estimate above --max-cost, an unknown ecosystem, or a resource that resolved fine but is "not enabled for generation" (the ids exist; the combination is not runnable β€” distinct from exit 4, which means "no such id") 2
--fail-on-substitution and the estimate reported a substituted checkpoint β€” nothing submitted (see Silent model substitution). πŸ”΄ A substitution that appears only on the submit reply is reported but exits 0: by then the charge has happened, and failing would strand a result you paid for. And against a server that does not report substitutions at all the flag is inert β€” exit 0, submitted, charged 1
--input that is malformed, declares a non-txt2img workflow, carries an envelope key (civitaiTip, …), or is combined with a content flag 2
No such --checkpoint / --lora version id 4
civitai workflows get / workflows cancel on an unknown workflow id (a read; spends nothing) 4
β€” πŸ”΄ submitted: the Buzz is already spent β€”
--timeout expired, or Ctrl-C while waiting β€” the job keeps running server-side and was not cancelled 1
The workflow finished failed / expired / canceled 1
The workflow succeeded but every output was filtered out (blocked / unavailable / hidden) 1

Upgrading

civitai upgrade replaces the running binary with the latest GitHub release:

civitai upgrade           # no-op (and says so) when already current
civitai upgrade --force   # reinstall anyway

The release is resolved from the public GitHub releases API β€” no token is ever sent β€” and the downloaded archive is verified against its SHA-256 entry in the release's checksums.txt before anything is replaced. A mismatch, or a release carrying no checksums.txt at all, aborts and leaves the current binary untouched: it will not upgrade without integrity verification.

If this binary came from Homebrew, upgrade does not self-replace it β€” it tells you to run brew upgrade civitai/tap/civitai, so the package manager keeps owning the file. --force overrides that and self-replaces anyway.

The other install paths update the way they normally do β€” npm install -g @civitai/cli@latest, nix profile upgrade, or re-running go install …@latest.

Separately from this command, the CLI runs a background check for a newer release and prints a one-line notice; --no-update-check (or CIVITAI_NO_UPDATE_CHECK=1) turns that off, which is what you want in CI.

Global flags

These are accepted by every command:

Flag What it does
-v, --version Print the version and exit. (civitai version prints version + commit + build date.)
-h, --help Help for any command. civitai --help also prints the exit-code contract.
--no-color Disable all colour and styling. Also via NO_COLOR or CIVITAI_NO_COLOR.
--color Force colour even when stdout is not a TTY. Also via CLICOLOR_FORCE or CIVITAI_COLOR.
--no-update-check Skip the background check for a newer release. Also via CIVITAI_NO_UPDATE_CHECK.

The colour contract, for pipelines. Colour is off by default whenever stdout is not a TTY, so a redirected or piped run already emits plain text with no escape sequences β€” you do not have to ask for anything. When you do want to override that, the precedence is fixed, highest first:

  1. --no-color / NO_COLOR / CIVITAI_NO_COLOR β†’ off
  2. --color / CLICOLOR_FORCE / CIVITAI_COLOR β†’ on
  3. otherwise: on if stdout is a TTY, off if it is not

Off always beats on, so a NO_COLOR in the environment cannot be re-enabled by a --color further down a pipeline. NO_COLOR follows the no-color.org convention β€” present and non-empty is what counts, not the value.

πŸ”΄ --json output is never styled, at any of those settings. It is written without passing through the presentation layer at all, so --json is always safe to pipe into jq regardless of how colour is configured or whether a TTY is attached.

Configuration

Setting Config key Env var Default
Personal API key token CIVITAI_TOKEN β€”
OAuth tokens (device login) auth_kind, access_token, refresh_token, token_expiry, scope β€” β€”
API base URL base_url CIVITAI_BASE_URL https://civitai.com
Submit endpoint β€” CIVITAI_SUBMIT_PATH /api/v1/blocks/submit-version
Skip the update check β€” CIVITAI_NO_UPDATE_CHECK unset (the check runs)
dev-tunnel SSH endpoint β€” CIVITAI_DEV_TUNNEL_ENDPOINT sish.civitai.com:2224
Disable colour β€” NO_COLOR, CIVITAI_NO_COLOR unset
Force colour β€” CLICOLOR_FORCE, CIVITAI_COLOR unset

Where a setting also has a flag β€” --token, --tunnel-endpoint, and the colour and update-check flags β€” the flag wins over the environment. See Global flags for the full colour precedence.

Config lives at ~/.config/civitai/config.yaml (honours XDG_CONFIG_HOME), written owner-readable only.

Exit codes

civitai returns a differentiated exit code so scripts can branch on the kind of failure without parsing stderr. The human-readable error message is unchanged by this β€” only echo $? differs.

The table is the index. Each code's full ledger β€” the paths it covers, the residuals it deliberately does not, and the rules a script has to branch on β€” is in that code's ### Exit code N subsection below. civitai --help prints the table's summaries and points here for the rest.

Code Meaning
0 Success.
1 Generic / unclassified error. A filesystem failure lands here, and so does a validation verdict β€” an invalid manifest, or a real directory holding no manifest. Detail
2 Usage error β€” a bad flag, a missing required flag or argument, a bad flag value, or a path that does not exist / is not a directory. Detail
3 Not authorized β€” login required, token invalid/expired, or the credential lacks the needed scope (HTTP 401/403). Detail
4 Not found β€” the requested resource does not exist. Detail
5 Network/transport failure or service unavailable β€” the code to retry on. Detail
6 Rate limited β€” throttled by the API (HTTP 429). 
# Branch on failure kind
if ! civitai models get "$id" >/dev/null 2>&1; then
  case $? in
    3) echo "log in first: civitai login" ;;
    4) echo "no such model: $id" ;;
    5|6) echo "transient β€” retry later" ;;
    *) echo "failed" ;;
  esac
fi
Exit code 1
  • A filesystem failure lands here β€” a file that exists but cannot be read, an unwritable config directory, an I/O error. It is neither a mistake about the invocation (2) nor a transport failure (5), and there is no filesystem-specific code.
  • A validation verdict lands here, and deliberately not on 2: civitai app validate exits 1 when the manifest is invalid, and likewise when the directory you named is a real directory with no block.manifest.json at its root β€” you pointed at a real place, so the invocation was right and the project is wrong. (A path that does not exist, or that is not a directory, is the invocation being wrong, and exits 2.)
  • When validation produces a result, civitai app validate --json prints it in full and its ok field is the structured form of the same answer; a failure that produces no result at all β€” a project directory the CLI cannot stat, say, because it is unreadable or because a path component below it is not a directory β€” still exits 1 with nothing on stdout, so branch on the exit code before parsing. The full exitβ†’stdout table is in The --json result shape.
  • A resource that exists but is not ready lands here too, and deliberately not on 4: civitai app metrics <slug> for an app whose submitted version is still in review exits 1, because the slug is right and the app does exist β€” only its analytics do not exist yet, and the error names civitai app status <slug> as the next command. 4 stays reserved for a slug with no submissions at all, so the two remain separately actionable: fix the slug, versus wait for approval.
Exit code 2
  • Usage error β€” a bad flag, a missing required flag or argument (e.g. civitai app withdraw with no publish-request id), a bad flag value (--limit out of range, a non-integer id, --template nope), or a request the API rejected as malformed (HTTP 400, e.g. a bad --period/--sort enum).
  • This does not depend on where the refusal happens: a mistake the CLI catches locally and one the server rejects both exit 2.
  • A local image the CLI refuses before uploading anything (civitai app listing set-icon <file>, civitai generate --image) exits 2 when the file is missing, empty, a directory, over the size cap, or not a PNG/JPEG/WebP β€” but a file that exists and cannot be read (permissions, an I/O error) is a filesystem failure rather than a mistake about the invocation, and exits 1, not 2.
  • That split is not images-only and it is not flags-only β€” it holds for a flag's value and a positional argument alike, over the paths listed here: civitai generate --input <file> likewise exits 2 for a path that is not there or is a directory, and 1 when the file is there and the read fails.
  • The project commands take a positional path and refuse it the same way: civitai app validate <dir> and civitai app submit <dir> exit 2 when the path does not exist or is not a directory, because both are mistakes about the invocation. A directory that does exist but holds no block.manifest.json is a validation verdict instead, and exits 1.
  • app listing set-cover and app listing add-screenshot take the same positional <file> and refuse it the same way. (The CLI has no --file image flag at all: the only --file is civitai download --file, which picks a file inside a model version.)
  • Paths outside that list are not covered, and mostly exit 1. civitai app listing … --dir <missing> exits 1 (it reports "no block.manifest.json found in …", the same way it does for a directory that is really there but holds no manifest), and so does civitai app submit … --out <path under a directory that does not exist>. Both are stated rather than promised: this is a ledger of the paths the split is published for, not a claim about every path in the CLI.
  • A usage error emits no JSON object, in every mode. civitai app validate /nope --json therefore writes nothing to stdout and exits 2; it used to print {"ok": false, …} and exit 1, which reported a nonexistent path as a validation result. Scripts that parsed that object must branch on the exit code first.
Exit code 3
  • Authentication/authorization β€” login required, token invalid/expired, or the credential lacks the needed scope (HTTP 401/403, or no token configured).
  • civitai generate refines this: several of its failures are not credential problems but would otherwise land here or on 2, so they exit 1 instead and a script never loops on civitai login. A muted account or incomplete onboarding arrives as a bare 403 that is byte-identical to a missing scope; out of Buzz and generation disabled arrive as 400 (the upstream 403 is re-thrown server-side as a tRPC BAD_REQUEST), which would otherwise read as "bad flags". See Generate.
Exit code 4
  • Usually an HTTP 404, but not always: some lookups answer 200 with an empty result set instead (civitai app status <slug> for an unregistered slug, civitai users get for an unknown username), and those exit 4 too.
  • The same question therefore exits the same way however the API happens to phrase the miss.
Exit code 5
  • Network/transport failure or service unavailable β€” dial/timeout, or HTTP 502/503/504 after retries.
  • This is the code to retry on, so a filesystem failure never lands here however retryable its errno looks: a permissions or I/O problem does not fix itself, and a loop that sleeps and re-runs would never terminate. Those exit 1.

Troubleshooting

Look up the message you got. Every row's left column is a fragment of a string this CLI really prints, so searching this page for a few words of your error should land you on the right row. (A test asserts that each of these strings still exists in the source, so the index cannot quietly go stale the way a hand-written list does.)

Credentials and access
You saw What it means Where to read more
no token configured Nothing is logged in. Run civitai login, or set CIVITAI_TOKEN. This includes civitai app list / app view: the App store is not an anonymous read. Submit & auth, Browse the App store
not logged in (401) The credential is present but invalid or expired. OAuth tokens refresh themselves; when the refresh token has also expired, log in again. For a personal key, mint a new one at civitai.com/user/account. Submit & auth
forbidden (403) Usually the invite-only Apps beta rather than a broken token β€” the same account reads the public API fine. Submit & auth
not permitted for your account (403) Managing a store listing needs Apps-author access, which is a narrower grant than being able to submit. Listing media requirements
not permitted to read this app's analytics (403) app metrics needs a full-scope personal API key. An OAuth login is refused here even when it can submit. App metrics
block lacks ai:write:budgeted scope Printed by your app at runtime under dev:live. The dev token was minted without --spend, so the CLI filtered the budgeted-spend scope out β€” it never requests that scope implicitly, even when your manifest declares it. Local dev loop
insufficient Buzz / generation disabled Not credential problems, which is why they exit 1 rather than 3 β€” a script must not loop on civitai login for either. Exit codes specific to generate
rate limited (429) Throttled; exit 6. For deep paging use --cursor rather than --page. Exit codes
Scaffolding a project
You saw What it means Where to read more
cannot derive a slug from / cannot appear in a blockId The name holds characters the blockId alphabet cannot carry, and dropping them would mint a different permanent public id than you typed. Choose one yourself with --slug <slug>. The blockId
is not valid UTF-8 The same refusal one step earlier: the name's bytes cannot be read at all. Pass --slug, and a --name that is valid UTF-8. The blockId
refusing to overwrite. Scaffold somewhere else app create / app init will not clobber a non-empty directory, and there is deliberately no --force β€” overwriting a directory you already have is not recoverable. Use --dir <new path>, or remove the directory first. Templates
Validating and submitting
You saw What it means Where to read more
is this an App project? There is no block.manifest.json at the path you named. The path itself was fine β€” which is why this exits 1 and not 2. Validate fidelity
no such directory β€” pass the path to an App project root The path does not exist. This is a usage error: exit 2, and --json prints nothing at all. Exit codes
is not a directory β€” pass the App project ROOT You pointed at a file β€” often the manifest itself. Pass the directory holding it. Exit 2. Exit codes
it did NOT check that the file is loaded The BLOCK_READY advisory on its weak tier: it could not resolve what your index.html loads, so it only checked whether some file mentions the message. The lines that follow name what it could not follow. The host handshake
nothing index.html loads reaches it The strong tier: the emitter is in your project but nothing the browser loads reaches it β€” an orphan file. Copying civitai-host.js in is only half the fix; it has to be referenced too. The host handshake
no lockfile is committed / is not a lockfile The platform build installs strictly from the committed lockfile, so a missing one β€” or a zero-byte one created with touch β€” fails the build server-side. A lockfile is generated by the package manager, never hand-written. Validate fidelity
refusing to submit without --yes A submit that would really upload asked for confirmation and found no TTY. Pass --yes in CI, or --package-only to just write the .zip. Command reference
Generating
You saw What it means Where to read more
refusing to spend Buzz without --yes The same gate on the money path. --dry-run prices the job without spending anything. Confirmation
--image requires --ecosystem Without an ecosystem the server never promotes the job to image-to-image: your images are silently dropped and you are billed for a plain text-to-image run. Hence a refusal rather than a warning. Image-to-image
interrupted while waiting The generation is still running and has already been charged. Ctrl-C stopped the wait, not the job. Re-attach with civitai workflows get <id>. Waiting, downloading, and re-attaching
model substituted The server ran a different checkpoint than you asked for and billed for what ran. Warned by default; --fail-on-substitution turns it into a refusal on the estimate, before any spend. Silent model substitution
Everything else
You saw What it means Where to read more
has no approved App Block yet The slug is right and the app exists β€” its analytics do not, because the version is still in review. Exit 1, not 4. App metrics
no such app for your account The slug matches none of your submissions. List them with civitai app status. Submission status
is ambiguous β€” it matches A model version has several files sharing that name. Select one by its numeric file id with --file <id>. Download model files
SHA256 mismatch for A download's hash did not match, and the partial file was deleted. Retry β€” this is integrity checking working, not a bug. Download model files
checksum mismatch for The same, during civitai upgrade. The binary was not replaced. Upgrading
git is required for `civitai app pull` app pull shells out to git, which is not on your PATH. Pull your app's repository

Still stuck? Every command takes --help, civitai --help prints the exit-code contract, and failures are differentiated by exit code β€” so a script can branch on the kind of failure without matching any of these strings.

Development

make ci      # go mod tidy + vet + test + build
make lint    # golangci-lint β€” a SEPARATE job; `make ci` does not run it
make test
make build   # -> bin/civitai
make fmt
go test ./... -cover

πŸ”΄ make ci is not a mirror of CI. It runs tidy + vet + test + build and does not run lint, which is its own CI job β€” so run make lint too before calling a change done. It errors out when golangci-lint is not on PATH rather than degrading to something weaker, which is what makes a clean run mean anything.

CI is eight jobs, not four steps. .github/workflows/ci.yml runs build-test (vet + gofmt -s -l . + test + build), lint, schema-drift, pins-vs-published, ready-ack-runtime, template-page-vite, template-page-money and scaffold-currency on every push to main and every PR. Several exist to catch drift between this repo and the platform — the vendored schema, the scaffold's npm pins, the block→host handshake — which a plain go test cannot see.

Running and gating are different questions, and fewer of those jobs gate a merge than run. The measured set of required status checks, and the instruction to re-measure rather than trust a written copy, live in AGENTS.md item 11 β€” deliberately in one place, because a second copy is how the original claim went stale. Notably lint reports without blocking, which is another reason to run it locally.

Releasing

Releases are built by goreleaser from a GitHub Actions workflow on a v* tag push:

git tag v0.1.0
git push origin v0.1.0

This cross-compiles for linux/darwin/windows Γ— amd64/arm64, stamps version/commit/date, and creates a draft GitHub Release with archives + checksums.txt. It also renders the Homebrew cask and attaches it to the release, but does not push it to the tap β€” see below. See AGENTS.md for the full process and the secrets it needs (HOMEBREW_TAP_GITHUB_TOKEN).

πŸ”΄ There are three publication channels, and clicking "Publish release" fires the other two. .github/workflows/release-npm.yml publishes the npm/ wrapper as @civitai/cli β€” the very first install option at the top of this README β€” and .github/workflows/release-homebrew.yml pushes the cask to civitai/homebrew-tap. Both trigger on release: [published]. So publishing the draft is not the last step of the GitHub release; it is also, in the same click, an npm publish and a Homebrew release. That matters because npm unpublish is restricted: a bad version is corrected by publishing another one, not by taking it back.

πŸ”΄ Nothing downstream may act on a tag alone, and the Homebrew channel used to. Until #308, the same goreleaser run that created the draft release also pushed the cask bump, so the cask named a version whose archives 404 for everyone until a human clicked "Publish release". Measured on 2026-08-09, with v0.1.91 tagged at 01:09Z and still a draft hours later:

cask Casks/civitai.rb          version "0.1.91"
GET .../download/v0.1.91/civitai_0.1.91_linux_amd64.tar.gz   404
GET .../download/v0.1.90/civitai_0.1.90_linux_amd64.tar.gz   200
npm @civitai/cli                                             0.1.90   (correct)

brew install civitai/tap/civitai failed for every user for ~2 hours. npm was correct throughout because it already waited for release: [published]; the Homebrew channel now waits for the same event, and tools/caskcheck asserts the invariant β€” the cask must never name a version that is not publicly downloadable β€” on every publish and once a day, over real unauthenticated HTTP.

πŸ”΄ A failing scheduled run notifies almost nobody, so the daily check files its own report. GitHub sends scheduled-workflow notifications only to the one account that last edited the cron line, through that person's own per-user Actions setting β€” there is no org-, team- or repo-level failure notification. Measured in this repo: every scheduled run of bump-flake-vendorhash.yml failed on 2026-07-20, 2026-07-27 and 2026-08-03, and no issue was ever filed about it. So release-homebrew.yml opens a GitHub issue titled [cask-check] … when the check is not green, rewrites that one issue's body on each subsequent failure (an edit notifies nobody, so a week of failures is one issue and zero comments), comments only when the kind of failure changes, and closes the issue when the check goes green β€” which is what makes the next failure notify again. It distinguishes "the check failed" from "the check could not run": an unreachable tap, an unreadable cask, an unclassifiable error and a verify job that died before producing a verdict each get their own wording, and none of them is allowed to read as a clean bill of health.

gh workflow run release-homebrew.yml -f drill=broken|lagging|unmeasurable is a fire drill: it points the check at a fixture, opens a real issue exactly as a real failure would, and cannot touch the tap. Run it after changing any of this β€” a notification path nobody has watched work is not a notification path.

A cask that merely lags a published release is green for the first 24h, because that is the normal state right after a publish and a permanently-red check is worse than none. After that it is a finding of its own, distinct from the 404 outage above and explicitly not claiming users are broken: the only thing allowed to move the cask is that release: published job, so a cask still lagging a day later means the event never reached it — a dropped webhook, an expired HOMEBREW_TAP_GITHUB_TOKEN, a failed push nobody read. The threshold is measured, not guessed: across the 30 most recent releases the gap from published_at to the tap commit is ~1 minute, and the tag→publish gap (a strictly larger window) has a median of ~2m30s and a worst case of 1h55m — the 2026-08-09 incident itself.

Authentication for that job is OIDC trusted publishing β€” there is no NPM_TOKEN secret. The trust is bound to the repository and to that workflow file's path, so moving or renaming release-npm.yml breaks publishing, and no secret rotation will fix it.

License

Apache License 2.0.

Documentation ΒΆ

Index ΒΆ

Constants ΒΆ

This section is empty.

Variables ΒΆ

View Source 
var ExamplesFS embed.FS

ExamplesFS holds the real example manifests (copied from the shipping civitai-block-* apps). They are embedded so the validate test can assert the "examples validate clean" claim against the same files the README points to.

View Source 
var SchemaJSON []byte

SchemaJSON is the vendored App manifest JSON Schema, embedded so the CLI validates against the same contract it ships. The file is the canonical copy intended to also be published server-side (see README).

Functions ΒΆ

This section is empty.

Types ΒΆ

This section is empty.

Source Files ΒΆ

View all Source files

Directories ΒΆ

Path Synopsis
cmd
civitai command
Command civitai is the unified Civitai CLI.
 Command civitai is the unified Civitai CLI.
internal
antipattern
Package antipattern is the scaffold-currency "born-broken app" gate.
 Package antipattern is the scaffold-currency "born-broken app" gate.
appapi
Package appapi is the CLI-internal client for the Civitai App Blocks developer surface: bundle submission, submission status / withdraw, the per-app dev-token and dev-tunnel machinery, the owner-only Forgejo clone info, and the OAuth device-authorization login flow.
 Package appapi is the CLI-internal client for the Civitai App Blocks developer surface: bundle submission, submission status / withdraw, the per-app dev-token and dev-tunnel machinery, the owner-only Forgejo clone info, and the OAuth device-authorization login flow.
auth
Package auth bridges the persisted config and the civitai.TokenSource contract: it yields a Bearer token for the api client, refreshing device-flow OAuth tokens (and persisting the rotated refresh token) when they expire or after a 401.
 Package auth bridges the persisted config and the civitai.TokenSource contract: it yields a Bearer token for the api client, refreshing device-flow OAuth tokens (and persisting the rotated refresh token) when they expire or after a 401.
blockproto
Package blockproto is the CLI's single authority for the block -> host postMessage contract that scaffolded, SDK-free apps must implement themselves.
 Package blockproto is the CLI's single authority for the block -> host postMessage contract that scaffolded, SDK-free apps must implement themselves.
cmd
Package cmd wires the cobra command tree for the civitai CLI.
 Package cmd wires the cobra command tree for the civitai CLI.
config
Package config handles the CLI's persisted configuration (API base URL and token) via Viper.
 Package config handles the CLI's persisted configuration (API base URL and token) via Viper.
devtunnel
Package devtunnel holds the transport-level pieces of `civitai app dev-tunnel`: the EPHEMERAL SSH keypair the CLI mints per session, the reverse-tunnel dialer (`ssh -R` to the sish endpoint) behind an interface, and a small clock/timer seam β€” so the command's lifecycle (mint β†’ tunnel β†’ teardown on signal / idle) is unit-testable without a live server or a real network.
 Package devtunnel holds the transport-level pieces of `civitai app dev-tunnel`: the EPHEMERAL SSH keypair the CLI mints per session, the reverse-tunnel dialer (`ssh -R` to the sish endpoint) behind an interface, and a small clock/timer seam β€” so the command's lifecycle (mint β†’ tunnel β†’ teardown on signal / idle) is unit-testable without a live server or a real network.
dnsprobe
Package dnsprobe resolves a hostname via DNS-over-HTTPS (DoH) to Cloudflare, bypassing the OS/local resolver, and builds an *http.Client that dials the DoH-resolved IP for that host.
 Package dnsprobe resolves a hostname via DNS-over-HTTPS (DoH) to Cloudflare, bypassing the OS/local resolver, and builds an *http.Client that dials the DoH-resolved IP for that host.
dogfoodguard command
Command dogfoodguard classifies a `civitai` invocation for the dogfood sandbox (scripts/dogfood-sandbox.sh) by asking the REAL command tree what the argv means, and prints a machine-readable verdict.
 Command dogfoodguard classifies a `civitai` invocation for the dogfood sandbox (scripts/dogfood-sandbox.sh) by asking the REAL command tree what the argv means, and prints a machine-readable verdict.
genapi
Package genapi is the CLI-internal client for the orchestrator generation graph (the tRPC `orchestrator.*` procedures behind `civitai generate`).
 Package genapi is the CLI-internal client for the orchestrator generation graph (the tRPC `orchestrator.*` procedures behind `civitai generate`).
manifest
Package manifest holds the App manifest filename constant and a lightweight reader for the fields the CLI needs (slug/version/name).
 Package manifest holds the App manifest filename constant and a lightweight reader for the fields the CLI needs (slug/version/name).
pkgzip
Package pkgzip packages an App project directory into the canonical ZIP the platform build recipe expects: the SOURCE tree (manifest + src + build config), with build artifacts and VCS/dependency dirs excluded.
 Package pkgzip packages an App project directory into the canonical ZIP the platform build recipe expects: the SOURCE tree (manifest + src + build config), with build artifacts and VCS/dependency dirs excluded.
scaffold
Shared @civitai/* scaffold-pin logic.
 Shared @civitai/* scaffold-pin logic.
scaffold/cmd/bump-pins command
Command bump-pins keeps the scaffold's @civitai/* pins in lockstep with npm.
 Command bump-pins keeps the scaffold's @civitai/* pins in lockstep with npm.
ui
Package ui is the single, cohesive presentation layer for the civitai CLI.
 Package ui is the single, cohesive presentation layer for the civitai CLI.
validate
Package validate checks a block.manifest.json against the vendored JSON Schema plus structural project checks (manifest at root, build coherence).
 Package validate checks a block.manifest.json against the vendored JSON Schema plus structural project checks (manifest at root, build coherence).
pkg
civitai
Package civitai is the importable HTTP client SDK for the Civitai API.
 Package civitai is the importable HTTP client SDK for the Civitai API.
tools
caskcheck command
Command caskcheck asserts the ONE invariant the Homebrew channel has to hold:
 Command caskcheck asserts the ONE invariant the Homebrew channel has to hold:

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL
go.dev uses cookies from Google to deliver and enhance the quality of its services and to analyze traffic. Learn more.