Aetox

README

Aetox

A Windows desktop app that finishes the work on your machine — files, browser, shell, documents.

ภาษาไทย · Website · Download · Architecture · Every decision, and why

---

What it is

Aetox is a desktop application for Windows that runs an AI agent against your own machine. You describe what needs doing; it reads and writes real files, runs real commands in a real shell, and drives a real browser you can watch.

It is one self-contained 41.5 MB executable. There is no runtime to install alongside it, no node_modules, no bundled copy of Chromium. It talks to whichever model you point it at — a hosted API, a subscription you already pay for, or a 9B running in LM Studio or Ollama on your own GPU — and the capability comes from the app rather than from the model's parameters. That is why a small local model can still read a picture, transcribe a recording, and produce a spreadsheet: image_ocr, audio_transcribe and sheet_write are the app's, not the model's.

Two concrete jobs, to make that less abstract. "Go through this folder of receipts and give me one spreadsheet" — it OCRs each image, works out the totals in a JavaScript interpreter compiled into the binary and shows you the script beside the answer, then writes a real .xlsx with live formulas. "Find why the login test is flaky" — it greps the repo, runs the suite in a terminal tab you can watch, reads the failure, and edits the file.

The interface ships in Thai and English, and Thai is the default. Every string exists in both; the language switch is in Settings and in the first-run wizard. This README is in English; ภาษาไทย is here.

Install

Windows 10 or later, x64. You do not need an API key to start — a built-in aetox provider ships five test models that exercise the real machinery (real tool calls, a real delegation to a sub-agent, a long reasoning stream), so you can see what the app does before signing up for anything.

Installer — aetox-amd64-installer.exe (15.8 MB)

It also fetches and SHA256-verifies the outside programs Aetox uses: WebView2, Tesseract (with Thai), poppler, ffmpeg, and a starter whisper model for offline speech. Any one of them failing skips that step with an explanation rather than aborting the install.

Scoop

scoop install https://raw.githubusercontent.com/Mike0165115321/Aetox/main/scoop/aetox.json

Portable — the zip, unpack, run aetox.exe. This is the only channel that can update itself in place.

If SmartScreen or your antivirus complains

The installer is not code-signed yet, so the first run shows "Windows protected your PC" with an unknown publisher — More info → Run anyway. Releases are signed: an ed25519 public key is compiled into the binary and the updater verifies the signature over checksums.txt before it trusts a single hash. An empty or wrong key refuses the update rather than falling back.

Linux and macOS

Not shipped. The engine and the desktop package both compile and their suites run under -race on Linux and macOS in CI; the browser pane is stubbed and packaging is not done.

1.0.0 is the Windows release. Until 2026-08-15 this line read "1.0.0 ships all three or it is not 1.0.0" — that criterion was changed by the owner, not met. Holding a stable Windows build behind a browser pane and an at-rest keystore that do not exist yet on the other two helps nobody already running it. Linux and macOS ship under the same bar, in a later release. See PLATFORM-SUPPORT.md for where the port actually stands, and DECISIONS §109 for why the bar moved.

Build it yourself

cd desktop
wails build          # → desktop/build/bin/aetox.exe
wails build -nsis    # with the installer

What you can do with it

Hand over a folder and get a file back. Point it at a directory of images, PDFs or recordings and ask for the thing you actually want. OCR (Thai and English), PDF text with the layout intact, and offline speech-to-text all feed the same conversation; sheet_write, doc_write and slides_write hand back a real Office file, built from Go's standard-library zip and XML with no third-party dependency — live formulas, dates that sort, Thai vowels in the right place.

Watch it work in a real browser. Not a headless scrape: a WebView2 window composited into the app, with an address bar, back/forward, DevTools, and eight device presets that resize the native window and zoom the page so CSS media queries genuinely fire. The agent opens, reads, clicks by reference and types into it while you watch the same tab.

Read what the model cannot see. image_ocr runs Tesseract with Thai and English, so a screenshot, a scan or a photographed form becomes text a 9B model can reason about — no vision model required, and the model that can see gets the image itself instead.

Point it at a repository. File tree, Monaco editor, unlimited real PTY terminal tabs, git, grep and glob over the whole tree, plus diagnostics and symbol backed by language servers the app installs on first use (gopls, typescript-language-server, svelteserver).

Delegate to a specialist. Type @doc, @sheet, @deck, @github, @automation or @research and your sentence reaches that agent word for word — not a paraphrase. Each is a folder on disk with its own prompt, its own memory, optionally its own model, and its own private skills.

Give it a job, not a step. Work that takes twenty moves is planned before it is worked, and todo_write puts that plan on screen while it runs, so what you watch is the order it chose rather than a spinner. Up to four specialists run at once: task hands work out and returns immediately, task_result collects it, so three jobs cost the time of the slowest rather than the sum. One that reaches a decision it should not make alone comes back as a question instead of a guess. One still working when your answer arrives keeps working — you collect it by the same id in a later turn, so the end of a turn is not a deadline. And before any of it runs there is a planning stance that can read anything and change nothing, on an allow-list, so a tool added next month is held back by default rather than slipping in.

Run on 2026-08-15, from one sentence — "find 20 CRMs a Thai SME could actually pick and give me a spreadsheet comparing them": 6m 51s, two agents, 42 tool calls between them — 8 by the assistant, 27 by research reading pricing pages, 7 by sheet — and one tool failure it worked around. The handoff between the two was the baton this README describes: research left a markdown report in the session's output folder and sheet was given the path, not the contents. Twenty rows came back, fourteen with a real numeric price sorted low to high, and six deliberately left blank with the reason written in beside them — quoted-only pricing, or a page that would not state a figure. Every row carries the date the page was read and the link the number came from, and says whether the page was opened in the browser or only searched. The blanks are the part worth trusting: a table with no gaps in it is a table that guessed.

Ask it about your own past work. Every conversation and every tool run lives in local SQLite with FTS5, so session_search across months of history is a query, not an inference — zero tokens, Thai and English alike.

Have it build automations in n8n or Windmill. Connect an instance you host and the automation agent lists, reads, creates and updates workflows in it, and can start the server for you from a command you saved. Read the honest limit before you rely on this.

Two doors, one app

One switch on the wordmark moves between Assistant ("Use, remember, and create") and Code ("Build, debug, and ship"). It is the same binary, the same data directory, the same settings, memory and permissions — switching doors is not switching apps, and the app remembers which one you were in. The door also scopes the chat list in SQL, so a run of coding sessions cannot starve the other list.

	Assistant	Code
Where it works	Your whole machine when no project is focused, or a project folder plus folders you add	The project folder you opened, plus folders you add
Rooms	Assistant · Projects · Agent team · Automation · Work	Code
The right-hand panel	Available	Available

The doors separate what the system carries, never what the AI is willing to do. The assistant has files and a shell and does software work with them; it does not hand a request back because it involves code.

The team

Six agents ship — deck, doc, sheet, github, automation, research — and hiring a seventh is dropping a folder into <DataRoot>/agents/. No release, no plugin API, no restart.

An agent's folder is its whole identity: AGENT.md (who it is, what desk it sits at, which tools it may narrow itself to, which model it pins), MEMORY.md (what it has learned), STARTERS.md (how an empty chat with it opens, per language), and a private skills/ folder no other agent can see.

That folder is also where the difference between a clever assistant and a company sits. Each agent pins its own model, so the one that opens twenty pricing pages can run on something cheap while the one that has to weigh what it found runs on something strong, and the bill follows the work instead of following the hardest task in it. Each keeps its own memory, so what the research agent learned about a source does not leak into the deck agent's judgement about slides. A single generalist has one model, one memory and one set of tools for every job it will ever be handed — and no way for you to add a seventh colleague to it.

You can delegate to one — the assistant calls task and up to four run concurrently — or you can talk to one directly, in a session bound to its tools, its memory and its prompt. @name from the composer is the third door: your sentence arrives verbatim, mention included, because a paraphrase is where the request goes wrong.

Agents never call each other. The star has one centre; multi-step work is a conveyor through the assistant, and the baton is a file path rather than the content. Separately, three sub-agents (explore, plan, general) are internal helpers — a fixed set, not extensible, deliberately.

What it learns, and what you approve

Aetox remembers across sessions, and nothing is written without you approving it.

The memory tool cannot write. It queues a proposal. Separately, a summarizer reads the tool-run log with no model call at all, clusters repeated failures by agent + tool + normalised error, and proposes a lesson once the same mistake has happened three times.

Everything lands in one review queue in Settings, and each card shows the body, the agent's stated reason, whose memory it would go into, and — for a replacement — the line it would overwrite. Approve or discard. What is kept is plain markdown you can open, edit line by line, or forget in place, and every decision is recorded permanently, so "why does it think that?" always has an answer. One switch turns the whole thing off.

It takes effect from the next session, not this one — a mid-conversation prompt change would invalidate the provider's prefix cache, which is the same reason the tool block never moves.

The other half is standing instructions: your own always-on markdown files that ride into every desk, every project and every agent. What you wrote, and what it worked out, are kept apart on purpose.

How it works, and what it can reach

A desk is the tool ceiling of a session. Three ship — assistant, coding, specialized — and a session's desk is fixed for its life. Desks are also what MCP servers and external connections are placed on, which is how a tool installed for one kind of work stays out of the others.

Where it may go. With a project focused, the workspace is that folder plus any folder you add — added folders get read and write with no prompt, the same rights as the root, because a second quieter tier would be a rule you never agreed to. With no project focused, the workspace is the machine, and writes land under output/<session>. One function resolves every path, symlinks and all, and there is deliberately no second check anywhere else.

Approval. Three levels, and one gate every tool call goes through — built-in tools, shell and MCP alike.

Level	What it asks about
Ask	Anything that is not a plain read inside the workspace
Unsafe only	Deletes, git, shell, and anything touching a path outside the workspace
Full access	Nothing. There is no carve-out.

MCP tools confirm at every level regardless, because their behaviour is defined by somebody else's server.

Shell commands are path-contained, not pattern-matched: a path hidden in a quoted argument, behind a flag, behind a redirect, or behind %VAR% / $VAR / ~ is still resolved and checked, and a command the scanner cannot read — $(...), backticks, -EncodedCommand, FromBase64String — is refused rather than guessed at. Every command run is appended to a 0600 audit log.

Refused to every file tool, in every mode: .ssh .aws .gnupg .azure .kube .netrc .git-credentials .config/gh .aetox, the Windows Credentials and Protect stores, Chrome / Edge / Firefox / Brave profiles, and Aetox's own credentials.json, oauth.json, mcp-servers.json and browser profile. Folder-picking refuses them too, so it fails at the door rather than as a confusing tool error later.

Your data.

	Where it stands
Chat history, tool runs, produced files	On your disk, in local SQLite and plain folders
Cutting the cloud off entirely	Run the model through LM Studio or Ollama — not one byte of the prompt leaves
API keys	Their own file, 0600, DPAPI-wrapped against your Windows account. Off Windows there is no encryption at rest — that is stated rather than implied
Secrets in logs	Stripped through one registry into all three sinks: debug log, shell audit log, and the buffer the bug-report form reads
MCP secrets	${env:VAR} indirection, so a key never lands in the settings file
Taking it with you	Export any chat to .md or .json, and import a .json back into any Aetox
Bug reports	The app transmits nothing. It prefills a GitHub issue, already scrubbed, and you read every line before sending it from your own account

There is no server of ours in the middle and no analytics. Using a cloud provider means that provider sees what its API normally sees, and nothing is routed through us.

Everything it can do

A tool count is not a reason to use anything, which is why this is down here.

40 tools reach the model on a fresh install; a default assistant session carries fewer, because a desk narrows the set. They cost about 9,844 tokens on every request before you have typed anything, against a ceiling of 10,100 that a test enforces — eight slots of headroom, argued for rather than spent.

Group	Tools
Files	read write edit apply_patch notebook_edit delete list glob grep
Running commands	shell (run · output · kill · list) git desk_terminal
Handing back files	doc_write sheet_write slides_write
Reading media	image_ocr video_ocr pdf_read audio_transcribe
Web and automation	browser (open · read · click · type) web_fetch web_search
Code work	diagnostics symbol github (repo_summary · search · read_file · list_files)
How the assistant works	task task_result task_answer ask_user todo_write suggest_task memory session_search calc time desk_list desk_open skills_list skill_view plugin_install

That table is generated from the registry the model is actually handed (go test ./desktop -run TestPrintReadmeToolTable -v), because a hand-kept list of what a program contains is a second source of truth for a question the program can answer — and this one drifted for months, still naming tools that had been folded into shell and github.

Connecting an automation engine adds one more packed tool — n8n (list · read · create · update · activate) or windmill (workspaces · list · read · create · update) — and nothing until then: a tool with no account behind it is withheld rather than shown and refused.

Growth goes where it costs nothing. A skill is a markdown document, not a tool: skills_list returns one line each and skill_view returns one body, so installing three hundred leaves the tool block exactly the same size. MCP servers are placed per desk and per agent, so a server added for video work is absent from an ordinary conversation — not hidden from the model, absent. Office writers reach only the specialized desk, so the assistant delegates for a .pptx rather than carrying three tools it rarely needs.

11 providers in the app — OpenAI · Anthropic · Gemini · DeepSeek · Qwen · Z.ai · OpenRouter · Codex · LM Studio · Ollama · and the built-in aetox. OpenRouter and Codex sign in; the rest take an API key or a local server address. The engine catalogue holds 18, and the remainder are not drawn in the window.

Local models are treated as first-class: Aetox asks LM Studio and Ollama which model is loaded rather than which exist, streams the answer and the reasoning, really calls tools, and counts tokens into the same statistics. You can switch provider or model mid-conversation and the full context follows — tool calls, tool results and compaction summaries, not just the visible text. One provider is active at a time; Aetox never silently reroutes your turn to a different paid one.

Automation: what it can and cannot do

Connect an n8n or Windmill instance you host and the automation agent can list, read, create, update, and — n8n only — activate workflows, and start your server from a command you saved.

It cannot run a workflow and see the result. There is no execution API call anywhere in this codebase; the closest thing is the agent clicking Execute in the vendor's own editor through the browser tool, which is not a verified run. Windmill has no activate either, so a flow it creates is saved and inert until you trigger it yourself. The agent says so out loud rather than implying otherwise, and a test exists whose only job is to keep it saying so.

There is no scheduler, and there will not be one. Aetox has no cloud, so a schedule would silently depend on your laptop never closing. n8n and Windmill are the clock; Aetox is the hands.

Measured, not claimed

The rules are in BENCHMARK.md, and its one standing rule is that a number which has not passed them may not appear here or on the website.

The dangerous number is the flattering one, because nobody audits a figure that makes them look good.

Aetox, measured 2026-08-13 on this build.

What you download	15.8 MB installer
What ends up on disk	41.5 MB, one file
Assembling a turn	0.32 ms · 174.9 KB allocated
Go tests	1,470 across 37 packages, 0 failures
Frontend tests	463 across 42 files, 0 failures
First launch (cold)	1.77 s ⁽ᵈ⁾
Every launch after	0.53 s ⁽ᵈ⁾
RAM committed	252 MB ⁽ᵈ⁾
Processes	7 ⁽ᵈ⁾

⁽ᵈ⁾ Measured 2026-07-27 on v0.9.2 under the rules, and not re-measured since. They are dated figures rather than current ones, and they are here rather than deleted because they did pass the rules on the day — which is the whole difference between an old number and a bad one.

Two things that number honestly. Assembling a turn was 0.12 ms and 96.2 KB when the block held 27 tools; it is 0.32 ms and 174.9 KB now that it holds more. That is a real regression, and it is still three ten-thousandths of a second — the time you wait is the model thinking. And the Go suite is green on Windows; CI on Linux and macOS is red, six failures, one of them a genuine sandbox hole rather than a bad test. Since 2026-08-15 those two jobs are reported rather than gating — Windows is what ships, and one shared verdict meant every Windows push went red for a port's unfinished edges until nobody read the colour at all. The failures are still on the run page and still to be fixed; what changed is that they no longer hide the platform that is done.

Against Zed, the harder ruler — native Rust, with a reputation for being light.

	Aetox	Zed
First launch (cold)	1.77 s	2.12 s
Every launch after	0.53 s	0.53 s
RAM committed	252 MB	471 MB
Disk	41.5 MB	419 MB

Both columns except Aetox's disk figure were measured 2026-07-27 on the same machine under the same rules, and neither has been re-measured — Zed is no longer installed here. A tie on warm launch with a native Rust editor is the result worth having; treat the row as dated rather than current.

The rest of this category ships 240 MB to 1 GB because an Electron app brings its own copy of Chromium. Aetox uses the WebView2 that Windows already has — and being straight about it, WebView2 is Chromium, so the memory it holds is not a win over Electron. The win is that you are not handed a second browser to store.

How these were measured, and what does not qualify

Disk — download the portable zip, unpack it, and read the size of the one aetox.exe inside: 43,514,880 bytes. Anyone can reproduce it in a minute. It replaces the 40.8 MB figure measured on v0.9.2 on 2026-08-07, which was correct then and is not now. Competitor sizes are measured after install from the install folder, never taken from a download page, and never from a folder holding user profiles or caches.

Launch, RAM and process count — bench.ps1 -Start, empty project, median of 5 runs after discarding the first, read after 60 seconds settled. A true cold launch needs a reboot first, because Windows keeps the app's files in its file cache afterwards.

Assembling a turn — bench.ps1 -Engine, median of 3 rounds.

What was removed from this section. An earlier version of this README published "97% of input tokens came from cache over six consecutive messages" and local first-token times of 1.42 s and 1.75 s. Neither has a source in this repository — no test, no log, no BENCHMARK entry — and the machine those local numbers describe did not have LM Studio installed. They are gone rather than date-stamped, because the rule above does not have an exception for numbers we would like to keep.

Status — v1.2.4

The core is in place. Release notes · roadmap · architecture.

Three things it does today that are worth knowing about:

• Undo, and answer variants. Every turn records the files it changed, so "Undo (3)" puts them back, and regenerating an answer tells you exactly which files it will restore first. Old answers are kept — arrows and "2 / 3" — and switching between them rewrites what the conversation continues from.
• The model draws. An SVG in an answer renders as a picture inside the reply, building itself shape by shape as it streams, with Copy and Save that bake the live theme's colours onto a clone and rasterise at 2×.
• You can talk into a running turn. Typing turns Stop back into Send and hands your sentence to the turn in flight; anything it could not fold in comes back as a queued bubble you can drop.

Next — agents working across turns rather than only inside one; a plan handed from the assistant door to the code door; a code-door team with defined roles.

Documentation

Architecture · Every decision, and why · What this company is · How a screen is designed · Benchmark rules · Platform support · Roadmap · Automation engines

Reporting bugs

Open an issue. The app has a door for this: Settings prefills a GitHub issue with your version, install channel and OS, folds the recent internal log into a <details> block with secrets already stripped, and hands it to you to read before you send it from your own account. Nothing is transmitted by the app.

Who makes this, and the licence

Aetox is written by one person. It exists because a model that can only produce text is half a tool, and the missing half — hands, permission, and a place to put the result — is an application problem rather than a model problem.

Code is under Apache-2.0: use, modify, redistribute and sell it freely, including in closed-source work. Two things asked in return — keep the NOTICE file and the copyright notice, and say which files you changed.

The name "Aetox" and the logo are trademarks and are not covered by the licence (Apache-2.0 clause 6). Fork freely; anything you redistribute carries your own name, so nobody is confused about which one is the original. Version 0.7.1 and earlier were MIT, and those versions stay MIT forever.

Aetox was not born to compete with anyone. It exists to stand where the market has a gap — not to be one more agent framework, and not to lock anyone into anything.

📧 phrmsawanachyphl@gmail.com · ❤️ Support the project

---

© 2026 Chayaphon Phromsawana · Apache-2.0 · "Aetox" is a trademark, not covered by the licence