lighthouse

README

Lighthouse

Lighthouse is a local-first store for coding-agent results and durable memory. Agents use one constrained Starlark program through the lighthouse code CLI. A Go server provides an optional, loopback-only dashboard for reviewing records, metrics, edits, audit events, and data transfer. The canonical store is a portable SQLite file; there is no MCP server or cloud service in the data path.

How an agent uses it

Install the lighthouse executable and the agent skill. Start substantive work with one context-load call, then reuse that context during the conversation. Batch any needed searches or reads into one script and emit only relevant results. After a write, consume the returned committed record, revision, and context patch so the active conversation can refresh its memory. Lighthouse cannot silently update hidden provider context.

lighthouse code --project my-project < examples/bootstrap.star

The code API includes load_memory, search, get, describe, audit, save, update, delete, upsert_memory, delete_memory, and emit. Documents accept flexible JSON custom fields without agent-controlled SQL schema changes. Deletions are explicit and audited. See Set up your coding agent and the examples.

Runtime, setup, and cost

The CLI and dashboard are Go programs. The code-call script language is Starlark, an embedded Python-like language interpreted in the already-running Go process. It is not a Python subprocess, and Lighthouse does not compile scripts on each call. The language has no shell, filesystem, environment, raw SQL, or network builtins. This avoids a runtime install and a per-call compilation step while keeping a typed, constrained API.

A local CLI invocation and SQLite read/write do not themselves consume model credits or call a hosted service. The agent still uses model tokens to plan the script and interpret the output; that usage is determined by the agent provider.

Option	Per-call setup	Fit for this workload
Embedded Go + Starlark (current)	Run the existing executable; interpret the short script	No separate interpreter or script compilation; typed capabilities
Python	Start an interpreter and load script	Easy for standalone scripts, but adds a runtime/process for every CLI call
Lua/LuaJIT	Start interpreter; possibly native dependency	Lightweight runtime, but another API/runtime to install and support
Go, Rust, C, C++	Compile source, or run a prebuilt executable	Fast native programs; per-request compilation is unnecessary overhead here
Java	Start JVM and provide a JRE, or package a native image	Larger setup for a local helper; not installed in the checked environment

These are architectural tradeoffs, not a portable performance benchmark: cold start and compilation vary by machine, toolchain, and packaging. The operation itself is a short local database access. The existing Go binary plus embedded Starlark is the simplest measured-to-implement choice; use a prebuilt Go executable for distribution rather than compiling a new helper on every agent turn.

Install the CLI

Install the published command without cloning the repository:

go install github.com/paulwritescode/lighthouse/cmd/lighthouse@latest

go install downloads the Lighthouse module and its dependencies, builds the lighthouse command for your machine, and places it in Go's binary directory. It does not install Go, configure a startup service, or start the dashboard. You do not need Git, a source checkout, Node.js, Python, a separate SQLite installation, or a database server.

Homebrew on macOS

On macOS, Homebrew can install the same tagged CLI release in one command from the Lighthouse tap:

brew install paulwritescode/lighthouse/lighthouse

Homebrew adds the tap and trusts this formula during installation, builds the pinned release with Go, and manages the installed executable. The shorter brew install lighthouse selects an unrelated Ethereum client from Homebrew Core. Use brew upgrade paulwritescode/lighthouse/lighthouse for later formula updates. If you installed a startup service, run lighthouse service restart after upgrading so it uses the new binary.

Prerequisites

• Go 1.27 or newer. Check with go version. If go is missing, install it from go.dev/dl and open a new terminal.
• Internet access for the initial install. Normal use of the installed CLI is local; Lighthouse does not send your database to a cloud service.
• Write access to Go's binary directory: GOBIN if set, otherwise $(go env GOPATH)/bin (usually ~/go/bin). No administrator privileges are needed for a per-user install.
• That directory on your PATH, so you can type lighthouse.

To pin a release for repeatable setups, replace @latest with a version:

go install github.com/paulwritescode/lighthouse/cmd/lighthouse@v0.8.1

The first time you run lighthouse, an interactive setup guides you through the local hostname and optional sign-in service one question at a time. System-level changes are opt-in. After setup, running lighthouse opens the control menu. Run lighthouse setup to revisit the setup questions.

If a startup service is already installed, restart it after a manual go install with lighthouse service restart. The built-in lighthouse update command installs into the directory of the running binary and handles stopping and restarting an installed service itself.

Put it on your PATH

Check whether your shell can find the command:

command -v lighthouse

If that prints nothing, add Go's binary directory to PATH. On macOS and Linux with zsh, add this line to ~/.zshrc (use your GOBIN value instead if you set one), then run source ~/.zshrc:

export PATH="$(go env GOPATH)/bin:$PATH"

For Bash, add the same line to ~/.bashrc or ~/.bash_profile. On Windows, add the directory printed by go env GOBIN (or go env GOPATH plus \bin) to your user Path, open a new PowerShell window, and check with Get-Command lighthouse.

Verify and initialize

lighthouse --help
lighthouse init

lighthouse init creates the default local database and prints its path. Change the location with LIGHTHOUSE_DB or --db /path/to/file.db. Keep the live database out of synced folders and active Git checkouts.

Install from source (contributors)

Only needed to modify Lighthouse or build the current checkout:

git clone https://github.com/paulwritescode/lighthouse.git
cd lighthouse
mkdir -p "$HOME/.local/bin"
go build -o "$HOME/.local/bin/lighthouse" ./cmd/lighthouse

Put ~/.local/bin on PATH, or run go install ./cmd/lighthouse from the repository root to use Go's standard binary directory.

Troubleshooting

• go: command not found: install Go 1.27 or newer and open a new terminal.
• The module cannot be downloaded: check internet access, and that any proxy or firewall allows your configured Go module proxy and source host.
• lighthouse not found after install: check go env GOBIN and go env GOPATH, add the effective directory to PATH, and open a new terminal.
• An older version runs: install the version you want explicitly, then use command -v lighthouse (or Get-Command lighthouse) to see which copy your shell runs.

go get adds Lighthouse as a dependency to a Go module; use go install to install the CLI executable.

Set up your coding agent

Installing the CLI is separate from installing agent instructions. Codex and Claude Code need the Lighthouse skill so they know when and how to call the CLI. Codex reads user skills from ~/.agents/skills and Claude Code from ~/.claude/skills.

If you installed only the CLI, download the versioned skill and profile files:

mkdir -p "$HOME/.agents/skills/lighthouse-memory" "$HOME/.claude/skills/lighthouse-memory" "$HOME/.codex" "$HOME/.claude"
curl -fsSL https://raw.githubusercontent.com/paulwritescode/lighthouse/v0.8.1/.agents/skills/lighthouse-memory/SKILL.md -o "$HOME/.agents/skills/lighthouse-memory/SKILL.md"
curl -fsSL https://raw.githubusercontent.com/paulwritescode/lighthouse/v0.8.1/.claude/skills/lighthouse-memory/SKILL.md -o "$HOME/.claude/skills/lighthouse-memory/SKILL.md"
curl -fsSL https://raw.githubusercontent.com/paulwritescode/lighthouse/v0.8.1/integrations/user-profile/codex-AGENTS.md >> "$HOME/.codex/AGENTS.md"
curl -fsSL https://raw.githubusercontent.com/paulwritescode/lighthouse/v0.8.1/integrations/user-profile/claude-CLAUDE.md >> "$HOME/.claude/CLAUDE.md"

Append the profile files once; if the destination already has instructions, they are kept and the new block is added at the end. If you cloned the repository, copy .agents/skills/lighthouse-memory and .claude/skills/lighthouse-memory into those directories and append the files in integrations/user-profile/ instead. For repository-only use, the project already contains both skill folders plus AGENTS.md and CLAUDE.md.

The integration needs the agent session to be able to start a local executable. It is not a remote connector. The CLI never sends data to an AI provider; the agent sees only the stdout it requested.

Use Lighthouse as a Go library

The module also exposes a public Go package for applications that want direct database CRUD without starting the HTTP dashboard or invoking a subprocess:

go get github.com/paulwritescode/lighthouse@v0.8.1

import (
    "context"
    "log"

    "github.com/paulwritescode/lighthouse"
)

func main() {
ctx := context.Background()
db, err := lighthouse.Open("./agent-results.db")
if err != nil {
    log.Fatal(err)
}
defer db.Close()

saved, err := db.CreateArtifact(ctx, lighthouse.ArtifactInput{
    Title: "Build result",
    Content: "The local build completed.",
    Project: "example-project",
    Agent: "codex",
    Kind: "result",
    Fields: map[string]any{"exit_code": 0},
})
if err != nil {
    log.Fatal(err)
}
}

The package exports artifact and memory create, read/search/list, update, and delete methods, context loading, counts, and audit access. It stores flexible document properties in JSON fields and owns its SQLite schema; callers should use the public API rather than writing SQL to Lighthouse tables. See the library package docs.

Use lighthouse.OpenDefault() to open the same database used by the CLI and dashboard, or pass an application-specific SQLite file to lighthouse.Open(path).

Run the dashboard

The optional dashboard binds to 127.0.0.1:8080; open http://127.0.0.1:8080 or http://lighthouse.localhost:8080. The CLI can also add a lighthouse entry to the local hosts file so you can use http://lighthouse:8080. It uses embedded Go templates, CSS, native forms, and a small local script for browser confirmation dialogs. No Node runtime, frontend build, React, or CSS framework is required to install or run Lighthouse.

The white dashboard uses a compact, flat interface for reviewing records. Overview shows record counts, seven-day activity, record mix, top tags, recent results, and the current SQLite footprint (database file plus active -wal file). Projects groups primary project keys and secondary tags across results and memory without a separate project entity; each project opens its associated results and memory. Activity holds safeguards and a paginated, expandable audit trail. Data tools holds import, export, and database deletion. The base SQLite file contains schema and indexes even when there are no records; the WAL can contain recent writes until checkpointing.

Each result and memory item can have one primary tag (the existing project key) and up to 20 secondary tags. Tags use lowercase kebab-case names, and the same name is reused across records. Results and Memory can filter by any tag and sort by title/key or primary tag; the Projects page opens each tag's grouped results and memory. Agent code calls accept tags=["research", "writing"] on save and upsert_memory, and lighthouse add accepts --tags research,writing. The Results and Memory screens support individual CRUD, manually selected batches, and all records matching the active filters. Bulk result edits can set project, kind, or agent; bulk memory edits can set project or kind. The Results screen links to batch import. Bulk deletes require confirmation; bulk writes are audited per record and capped at 10,000 records. The tables show the newest 50 results and 100 memory items; “select every matching” applies the filters to the full set up to the cap.

Deleting the database from Data tools requires typing DELETE, browser confirmation, and a native macOS administrator password dialog. The password is never sent to the dashboard. Lighthouse removes the database and journal files and creates a fresh, empty database for the still-running dashboard. Dashboard mutations require a same-origin form and CSRF token, and changes are audited. The CSRF key is stored with user-only permissions so restarting the dashboard during an update does not invalidate forms in open tabs. If a form is rejected, refresh that page and submit again. Agent CLI operations work while the dashboard is stopped.

Startup service and local hostname

Run lighthouse to open the interactive control menu, or use the commands directly:

lighthouse service install
lighthouse service status
lighthouse service start
lighthouse service stop
lighthouse service restart
lighthouse service force-stop
lighthouse service remove

Install asks for confirmation, creates a per-user service that starts at sign-in, and starts it immediately. Remove asks for confirmation, stops Lighthouse, and removes the startup entry. The service uses the current database path and loopback-only dashboard. macOS uses a LaunchAgent in ~/Library/LaunchAgents; Linux uses a systemd user unit in ~/.config/systemd/user; Windows creates a Task Scheduler logon task for the current user. Linux installations need the user's systemd session available at sign-in. No administrator permission is needed for startup services.

force-stop confirms before terminating the managed dashboard. macOS sends SIGKILL; systemd escalates a timed-out stop to SIGKILL; Windows ends the scheduled task instance. The service can be started again with lighthouse service start.

Register the short local URL with a separate confirmation:

lighthouse host register
lighthouse host status
lighthouse host remove

Registration adds only a marked Lighthouse block to the OS hosts file. macOS and Linux ask for the administrator password through sudo; Windows shows its standard elevation prompt. Removal deletes only that marked block. You can always use http://127.0.0.1:8080 or http://lighthouse.localhost:8080 without registering a hosts entry.

lighthouse update installs the latest published release. When a startup service is installed, the CLI stops it before replacing the binary and starts it again afterward. If installation fails, it attempts to restart the previous version. Updating while a dashboard is running manually leaves that foreground process in place; stop it with Ctrl+C and start it again to load the new binary.

The database defaults to the OS user configuration directory. Set LIGHTHOUSE_DB or pass --db /path/to/file.db to choose another file. Avoid placing a live database in a synced vault or active Git checkout.

Import, export, and storage

The dashboard and CLI support JSON, Markdown, YAML, TOML, PDF, and SQLite .db transfer. Portable data imports merge records and skip duplicate IDs or memory keys. A .db export is a consistent SQLite snapshot; importing a .db merges artifacts and active memory rather than replacing the live database. PDF export is a readable report; PDF imports preserve bounded raw source bytes without extracting text.

Lighthouse uses SQLite through the CGo-free modernc.org/sqlite driver, with FTS5 and an optional bundled vector extension. Keyword search is available locally. Semantic search requires embeddings to be supplied; Lighthouse does not silently send data to an embedding provider.

Security boundaries and known limits

• The dashboard binds to loopback and validates allowed hostnames. Mutations use CSRF protection and audit events.
• Starlark has bounded input, output, execution steps, and wall time, and exposes typed operations only. It does not have a hard memory limit; use trusted local scripts.
• Local storage has no built-in encryption or multi-user authentication. Protect the user account and database file accordingly.
• Semantic search requires embeddings to be supplied by the caller; Lighthouse does not generate or upload them automatically.

Documentation

The full documentation (installation, quickstart, agent setup, the code API, the dashboard, import and export, architecture, security, and why Lighthouse uses a code call instead of MCP) lives on the Lighthouse website. Its source is in paulwritescode/lighthouse-website.

Release history is in CHANGELOG.md. Runnable script samples are in examples/.

Documentation

Overview

Package lighthouse exposes the local Lighthouse database to Go programs.

Open a database path once, use Database methods with a context, and close it when finished. Applications can import this package directly or install the CLI separately with `go install github.com/paulwritescode/lighthouse/cmd/lighthouse@latest`.

Index

• func DefaultDatabasePath() (string, error)
• type Artifact
• type ArtifactInput
• type ChangeEvent
• type ContextPack
• type Counts
• type Database
  • func Open(path string) (*Database, error)
  • func OpenDefault() (*Database, error)
  • func (db *Database) Audit(ctx context.Context, limit int) ([]ChangeEvent, error)
  • func (db *Database) Close() error
  • func (db *Database) Counts(ctx context.Context) (Counts, error)
  • func (db *Database) CreateArtifact(ctx context.Context, input ArtifactInput) (Artifact, error)
  • func (db *Database) CreateMemory(ctx context.Context, input MemoryInput) (Memory, error)
  • func (db *Database) DeleteArtifact(ctx context.Context, id string) error
  • func (db *Database) DeleteMemory(ctx context.Context, id string) error
  • func (db *Database) GetArtifact(ctx context.Context, id string) (Artifact, error)
  • func (db *Database) GetMemory(ctx context.Context, id string) (Memory, error)
  • func (db *Database) ListArtifacts(ctx context.Context, project string, limit int) ([]Artifact, error)
  • func (db *Database) ListMemory(ctx context.Context, project string, limit int) ([]Memory, error)
  • func (db *Database) LoadMemory(ctx context.Context, session, project string, limit int) (ContextPack, error)
  • func (db *Database) SearchArtifacts(ctx context.Context, query, project string, limit int) ([]Artifact, error)
  • func (db *Database) UpdateArtifact(ctx context.Context, id string, fields map[string]any) (Artifact, error)
  • func (db *Database) UpdateMemory(ctx context.Context, id string, input MemoryInput) error
• type Memory
• type MemoryInput

Functions

func DefaultDatabasePath

func DefaultDatabasePath() (string, error)

DefaultDatabasePath returns the path used by the Lighthouse CLI, honoring LIGHTHOUSE_DB before the operating-system user configuration directory.

Types

type Artifact

type Artifact = store.Artifact

Artifact is one indexed Lighthouse document.

type ArtifactInput

type ArtifactInput struct {
	Title   string
	Content string
	Project string
	Agent   string
	Kind    string
	Fields  map[string]any
	Tags    []string
}

ArtifactInput contains the indexed fields and optional per-record JSON fields.

type ChangeEvent

type ChangeEvent = store.ChangeEvent

ChangeEvent is one audited data mutation.

type ContextPack

type ContextPack = store.ContextPack

ContextPack contains the context and sources loaded for a project or session.

type Counts

type Counts = store.Counts

Counts summarizes stored results, memories, projects, and audit activity.

type Database

type Database struct {
	// contains filtered or unexported fields
}

Database is a handle to one local Lighthouse SQLite database.

func Open

func Open(path string) (*Database, error)

Open opens or creates a Lighthouse SQLite database at path and applies the current Lighthouse migrations. Use ":memory:" for an ephemeral database.

func OpenDefault

func OpenDefault() (*Database, error)

OpenDefault opens the same database file used by the CLI by default.

func (*Database) Audit

func (db *Database) Audit(ctx context.Context, limit int) ([]ChangeEvent, error)

Audit returns the latest recorded changes.

func (*Database) Close

func (db *Database) Close() error

Close releases the underlying SQLite connection.

func (*Database) Counts

func (db *Database) Counts(ctx context.Context) (Counts, error)

Counts returns database record and change totals.

func (*Database) CreateArtifact

func (db *Database) CreateArtifact(ctx context.Context, input ArtifactInput) (Artifact, error)

CreateArtifact creates an indexed result with optional JSON-compatible fields.

func (*Database) CreateMemory

func (db *Database) CreateMemory(ctx context.Context, input MemoryInput) (Memory, error)

CreateMemory creates or updates a memory item with the same project/key.

func (*Database) DeleteArtifact

func (db *Database) DeleteArtifact(ctx context.Context, id string) error

DeleteArtifact removes an artifact and records an audit event.

func (*Database) DeleteMemory

func (db *Database) DeleteMemory(ctx context.Context, id string) error

DeleteMemory retracts a memory item and records an audit event.

func (*Database) GetArtifact

func (db *Database) GetArtifact(ctx context.Context, id string) (Artifact, error)

GetArtifact returns an artifact by its stable ID.

func (*Database) GetMemory

func (db *Database) GetMemory(ctx context.Context, id string) (Memory, error)

GetMemory returns an active memory item by its stable ID.

func (*Database) ListArtifacts

func (db *Database) ListArtifacts(ctx context.Context, project string, limit int) ([]Artifact, error)

ListArtifacts returns the most recently created artifacts in the project scope.

func (*Database) ListMemory

func (db *Database) ListMemory(ctx context.Context, project string, limit int) ([]Memory, error)

ListMemory returns active memory items in the requested project scope.

func (*Database) LoadMemory

func (db *Database) LoadMemory(ctx context.Context, session, project string, limit int) (ContextPack, error)

LoadMemory creates a context snapshot for the session and project.

func (*Database) SearchArtifacts

func (db *Database) SearchArtifacts(ctx context.Context, query, project string, limit int) ([]Artifact, error)

SearchArtifacts finds artifacts by full text and optional project scope.

func (*Database) UpdateArtifact

func (db *Database) UpdateArtifact(ctx context.Context, id string, fields map[string]any) (Artifact, error)

UpdateArtifact applies indexed attributes and merges JSON fields by key.

func (*Database) UpdateMemory

func (db *Database) UpdateMemory(ctx context.Context, id string, input MemoryInput) error

UpdateMemory replaces the selected memory item's key, body, scope, and metadata.

type Memory

type Memory = store.Memory

Memory is a durable, revisioned Lighthouse context item.

type MemoryInput

type MemoryInput struct {
	Key      string
	Body     string
	Project  string
	Kind     string
	SourceID string
	Tags     []string
}

MemoryInput contains a durable, project-scoped fact or preference.