<div align="center">

██╗   ██╗ ██████╗ ██╗  ██╗ █████╗ ██╗    ██████╗ ██████╗ ██████╗ ███████╗██████╗
██║   ██║██╔═══██╗╚██╗██╔╝██╔══██╗██║   ██╔════╝██╔═══██╗██╔══██╗██╔════╝██╔══██╗
██║   ██║██║   ██║ ╚███╔╝ ███████║██║   ██║     ██║   ██║██║  ██║█████╗  ██████╔╝
╚██╗ ██╔╝██║   ██║ ██╔██╗ ██╔══██║██║   ██║     ██║   ██║██║  ██║██╔══╝  ██╔══██╗
 ╚████╔╝ ╚██████╔╝██╔╝ ██╗██║  ██║██║   ╚██████╗╚██████╔╝██████╔╝███████╗██║  ██║
  ╚═══╝   ╚═════╝ ╚═╝  ╚═╝╚═╝  ╚═╝╚═╝    ╚═════╝ ╚═════╝ ╚═════╝ ╚══════╝╚═╝  ╚═╝

A desktop coding agent for Windows and Linux.<br>
Any model on OpenRouter, or one running on your own PC. You pay only for what you use.

[ v0.2.87 ] [ Windows · Linux ] [ MIT ]

Download · What it does · Build from source · Layout

</div>


 ┌──────────┐   ask    ┌──────────────────┐   any model    ┌──────────────────────┐
 │   you    │ ───────▶ │    agent loop    │ ─────────────▶ │ OpenRouter  or  your │
 └──────────┘          │  read · edit ·   │ ◀───────────── │ own PC (llama.cpp)   │
      ▲                │  run · check     │   tool calls   └──────────────────────┘
      │    the result  └────────┬─────────┘
      └─────────────────────────┘   files · shell · preview · tests

You describe what you want. VoxAI Coder reads your project, edits files, runs commands,
checks its work in a browser and reports back. The agent loop is its own (Node, inside
Electron): no external agent binary and no subscription, just your
OpenRouter key or a local model.

▓▒░ Download

 ╔═══════════╦══════════════════════════════════════════════════════╗
 ║  Windows  ║  VoxAI-Coder-Setup-0.2.87.exe   (installer, 115 MB)  ║
 ║           ║  VoxAI-Coder-0.2.87-portable.exe   (no install)      ║
 ╠═══════════╬══════════════════════════════════════════════════════╣
 ║  Linux    ║  installer script + AppImage, when the Linux build   ║
 ║           ║  for the current version is out                      ║
 ╚═══════════╩══════════════════════════════════════════════════════╝

• Windows: the installer from CodeRook, or both files from GitHub Releases.
• Linux: from the Releases tab on CodeRook. Installing, what differs on Linux, and fixes for common problems: [the Linux guide](docs/LINUX.md).
• macOS: no build yet. The code is ready for one; [the macOS notes](docs/MACOS.md) say what is done and what a port still needs.

The installed app keeps itself up to date. It checks for new versions when it opens,
downloads them in the background and offers to restart when one is ready. Updates come from
CodeRook and are installed only when their signature and checksum match. The portable exe
does not update itself.

── First run

 1 ▸ Settings ▸ Account     paste your OpenRouter key, Save
                            (a management key is optional: balance and adding credits)
 2 ▸ the chat bar           pick a model and a project folder
 3 ▸ type what you want     and send

Keys are kept in the system keychain (Electron safeStorage) and are sent only to
OpenRouter.

▓▒░ What it does

── Builds it, then checks it


• Reads, writes, edits and patches files, and searches the project. Shell commands run in a terminal panel you can watch.
• A preview browser beside the chat for web apps and games, and computer use for desktop apps.
• Permission modes: Ask first, Edit files, Edit and test, Auto (a safety check decides what needs asking) and Full access. Work stays inside the project folder.
• Plan mode, a running task list, and Review changes, then commit or open a pull request.
• Fork a session from any of your messages to try another way, keeping the first.
• TypeScript errors in a file come back with each edit (the project's own compiler, incremental), so the agent fixes them as it goes.

 /goal    keep working until a separate check agrees it is done
 /review  review the changes that are not committed yet
 /init    study the project and write AGENTS.md
 /rewind  undo the file changes from the last run
 /loop    run a task again every so often

── Any model, priced before you send

<img src="docs/screenshots/model-picker.png" alt="The model picker, browsed by company, with prices per million tokens" width="340">

• Every model on OpenRouter, browsed by company, with prices shown before you pick one.
• Optional Anthropic and OpenAI keys: Claude and OpenAI's own models then go straight to their maker, billed to that key, with Anthropic's own caching and thinking.
• Team mode: a stronger model plans and reviews, and a cheaper one writes the code.
• Sub-agents for side tasks, each with its own model.

── Models on your own PC

<img src="docs/screenshots/local-models.png" alt="Local models: search Hugging Face and download a GGUF model" width="640">

• Local models on Windows and Linux through llama.cpp (Vulkan). The engine downloads once and is checked against its published SHA-256.
• Search Hugging Face and download GGUF models inside the app; the downloads resume and are checked when they finish.

── Apps, skills and memory

<img src="docs/screenshots/apps-mcp.png" alt="Apps (MCP): add a browser, Windows apps, Blender, Figma or your own server" width="640">

• MCP apps: a browser, Windows apps, Blender, Figma, files or your own server. Add, test, switch on and off, or import the ones you set up for Claude Desktop or Cursor.
• Skills, built in (GitHub, GitLab, Codeberg, CodeRook, the preview browser and computer use) or your own.
• Project memory and AGENTS.md, so a new session starts knowing the project.

<img src="docs/screenshots/skills.png" alt="Skills: built-in skills and your own" width="640">

── Costs you can see, and fewer tokens spent

<img src="docs/screenshots/spending.png" alt="Spending: a limit per run, a limit per goal, and the balance" width="640">

• The cost of each message and session, a usage history with sortable columns, a limit per run and per goal, and your balance with a button to add credits.
• Fewer tokens on every model:

 ▸ one fixed session prompt, so the provider's cache keeps hitting
 ▸ only the memory that matters to the request
 ▸ a project overview and the project's own commands, up front
 ▸ old tool output pruned in batches, without breaking the cache
 ▸ a stop when the model goes in circles
 ▸ a check before the agent says it is done
 ▸ duplicate tools left out of what the model is sent

<img src="docs/screenshots/context.png" alt="Context: how much of the session each request carries, and when older turns are summarised" width="640">

▓▒░ Build from source

You need Node.js 20 or newer (developed on Node 24) and Windows or Linux.

cd voxai-coder
npm install
npm start        # run the app
npm test         # offline regression tests: no model calls, no credits
npm run dist     # Windows installer and portable exe in voxai-coder/release

The Linux build is npm run dist:linux; see [LINUX-BUILD.md](LINUX-BUILD.md). A Mac build is
npm run dist:mac (untested; see [docs/MACOS.md](docs/MACOS.md)). Working on
the code: [voxai-coder/README.md](voxai-coder/README.md).

▓▒░ Layout

VoxAI_Coder/
├── voxai-coder/            the app
│   ├── src/main/           Electron main: settings, keychain, sessions, updater,
│   │                       local engine, Hugging Face downloads
│   ├── src/backend/        agent loop, providers, tools, Team mode, cost tracking,
│   │                       prompt caching and pruning
│   ├── src/renderer/       the UI (a single index.html)
│   ├── src/preload/        the bridge between the UI and main (window.voxai)
│   ├── resources/          icons, built-in skills, the public update key
│   └── test/               regression tests
├── docs/LINUX.md           using VoxAI Coder on Linux, and fixing it
├── docs/MACOS.md           porting notes for macOS
├── docs/screenshots/       the pictures on this page
├── LINUX-BUILD.md          building on Linux and adding it to a release
└── benchmark/              runs against other agents

▓▒░ License

 MIT  ·  Copyright (c) 2026 Batty 251

[MIT](LICENSE): the code, the screenshots in docs/screenshots and everything else in this
repository. The balance in the screenshots is a sample.