taucode
Your models. Your workspace.
A desktop app for AI coding, conversations, and document work.
Bring your own API keys or local models, then choose the engine that runs them.
Documentation · Getting started · Builds & limitations · FAQ · Report a problem
taucode puts the model, agent engine, and workspace under your control. Use a cloud API, a local model server, or your own compatible gateway. Keep project conversations, file changes, terminal sessions, browser tasks, and agent activity together instead of moving between separate tools.
It builds on CodexMonitor, with a provider registry, multiple engines, dedicated Chat and Work modes, bounded agent teams, checkpoints, browser and computer tools, and a customizable desktop interface.
Early development: macOS has the broadest live verification. Windows has an unsigned x64 test installer and build/startup checks; Linux and iOS need further validation. There is no stable release guarantee or feature parity across platforms. Read the current status before installing.
The actual interface rendered by the offline UI preview with sample projects and messages. Screenshots do not show private workspaces or live provider sessions.
Three ways to work
| Mode | Use it for | What you get |
|---|
| Code | Working in a project or repository | Project conversations, model and engine selection, diffs, Git/worktrees, terminals, side chats, agents, and checkpoints. |
| Chat | Asking questions and discussing attached material | A separate conversation history, streaming replies, text/image attachments, Markdown export, and an optional restricted Web toggle. |
| Work | Turning a folder and a brief into deliverables | Spaces, queued tasks, plans, approvals, output files, and bundled document, spreadsheet, presentation, PDF, and chart skills. |
Code uses the selected enabled engine. Chat and Work use dedicated pi profiles; changing the Code engine does not change those profiles.
Chat
Work
Both screenshots use the offline preview's sample data.
Get started
New to AI tools? Start with the beginner guide, follow installation and first run, then try the first Work task. The glossary explains unfamiliar terms.
Download a test build
The October 30 test release includes:
| Your computer | File |
|---|
| Apple Silicon Mac | taucode_0.7.68_aarch64.dmg — development signed, not notarized. |
| Intel/AMD Windows x64 | taucode_0.7.68_x64-setup.exe — unsigned. |
These are experimental packages. Follow the step-by-step installation guide and read the build notes and limits. The download includes supporting runtimes; development tools are needed for source builds, not normal use of the packaged app.
Build from source (developers)
You'll need Node.js 22.19 or newer with npm, the stable Rust toolchain, Git, CMake, and your platform's Tauri prerequisites. Native dependencies require additional platform tools; Windows needs MSVC and LLVM/Clang, and macOS computer-use components need Xcode/Swift tooling.
git clone https://github.com/aiden0rchad/taucode.git
cd taucode/apps/taucode
npm ci
CARGO_NET_OFFLINE=false cargo fetch --locked --manifest-path src-tauri/Cargo.toml
npm run tauri:dev
The native app's Cargo configuration defaults to offline builds. The explicit network override above populates the dependency cache on a fresh checkout.
Windows source builds: use npm run tauri:dev:win for the platform configuration. Compilation and startup checks are narrower than full feature verification.
Just exploring the interface? Run npm run preview:ui from apps/taucode and open http://localhost:1421. This uses sample data and makes no provider requests. It cannot run real tasks or access native desktop features.
Connect a model
- Open Settings → Providers and add a cloud API or local server.
- Configure its authentication and choose Test & fetch models.
- Enable a model and choose your defaults under Settings → Agents.
- Check Settings → Engines and install the pi runtime if needed.
- Add a project in Code, start a Chat, or choose a folder in Work.
The setup checklist walks through appearance, provider, engine, project, features, and permissions. Reopen it from Settings → About → Setup checklist.
taucode uses provider credentials. It does not connect ChatGPT, Claude, or other consumer subscription accounts. API usage and provider limits apply separately. See providers and authentication.
What you can do
| Capability | Details |
|---|
| Choose providers | OpenAI, Anthropic, Google, OpenRouter, Groq, Mistral, xAI, DeepSeek, Z.ai, Ollama, LM Studio, and compatible endpoints. Compatibility depends on the model and engine. |
| Choose engines | pi by default, oh-my-pi, optional Codex, Claude Code through ACP, and compatible custom ACP/pi-RPC executables. |
| Work across projects | Group projects, keep conversation trees, fork side chats, inspect diffs, manage Git branches/worktrees, and use GitHub integrations through gh. |
| Delegate deliberately | Single model, custom delegation, or Code's Chain of Command, with role/model selection and concurrency, depth, time, and reported-token limits. |
| Connect tools | Configure stdio or HTTP MCP servers with per-server approval and tool filters. CodeGraph is an optional code-intelligence integration. |
| Use the browser | A native macOS browser pane, a Windows native pane with separate Chromium automation, and an advanced Chromium backend. |
| Use desktop apps | Experimental macOS computer use with explicit permissions, app restrictions, approval controls, and Stop/takeover behavior. |
| Recover and compare | Checkpoints and selective rewind, interrupted-turn recovery, an opt-in diagnostics journal, and versioned model/engine evaluations. |
| Make it yours | Light/dark/system appearance, built-in and custom themes, editable tokens, shape/border choices, and independent feature switches. |
Documentation
The documentation site has local search, a persistent guide sidebar, light/dark themes, and page outlines. Every guide is also readable directly in this repository.
Before you hand it real work
- Models receive the context you send. Provider choice determines where prompts, attachments, tool results, and screenshots are processed. A local app is not automatically an offline system.
- Review the access policy. Shell commands, connectors, custom engines, and computer use can have real effects. A trusted custom executable is not confined by an OS sandbox just because it speaks ACP.
- Keep backups. Checkpoints have exclusions and size limits. Rewind changes files and can delete files added after a checkpoint. It cannot undo remote actions, sent messages, or changes inside another app.
- Telemetry and diagnostics are separate. The frontend currently includes inherited Sentry reporting; the local diagnostics journal is opt-in. See privacy, including how to disable Sentry in a source build.
- Treat usage limits as reported usage. Missing prices stay unknown. Requests already in flight can exceed a reported-token threshold; these controls are not a guaranteed billing cap.
Development
The app lives in apps/taucode; the documentation site has its own package in docs.
# From apps/taucode
npm run typecheck
npm test
npm run build
# From the repository root
npm --prefix docs ci
npm --prefix docs run build
See development for Rust checks, the UI preview, documentation authoring, and the source map. The engineering reference links the original specifications and verification records separately from the user guides.
Bugs and contributions
Report a bug with the platform, build, engine, model/provider type, and steps to reproduce. Remove API keys, personal files, and private workspace details from attachments.
Read CONTRIBUTING.md before opening a pull request. Report vulnerabilities privately using SECURITY.md.
License and credits
MIT. taucode is derived from CodexMonitor by Thomas Ricouard; the original copyright and license are preserved. Embedded engines, dependencies, models, and app logos retain their own licenses. See third-party notices.
