A minimal, focused interface for Pi.
Download · macOS (Apple Silicon)
- Project pane with chat persistance
- Remote projects via SSH
- Support for Pi Extensions
- Archived sessions
- Mermaid diagram rendering
bun install
bun tauri devRequires Node.js 22.18+ and Pi on PATH or TAU_PI_PATH. Each checkout/worktree starts its own
Vite server on an available port and passes that URL to Tauri; HMR uses the
same port. Use bun tauri dev (not bunx tauri dev) so the launcher manages
the server and URL together. Other commands, including bun tauri build,
pass through to the Tauri CLI unchanged.
Every native dev process starts with disposable storage seeded with two sample
projects (atlas and notes) and five sessions. Pi still runs normally with your
existing auth, models, and extensions, but local transcripts and all Tau-owned
state belong to that process. Closing the app discards them; native restarts
(including Rust rebuilds) start fresh. Frontend HMR does not reset native storage.
Production builds always retain the existing persistent store.
bun run dev opens the same seed in a browser-only, in-memory playground with
local mock replies; each tab/reload starts fresh and needs no Pi installation.
bun run repro -- <scenario> keeps its explicit deterministic scenarios. Both
commands also use an available port. Playwright continues to use fixed port 1420.
See development storage for lifetime and isolation
boundaries.
- Search
- Steering
- File operations (explorer/editor)
- Keyboard shortcuts
- Command palette
- Support for other harnesses
- Integrated terminal and browser
Pi is good because it's lean, stable, and extensible. Instead of forcing opinions on you, it provides a solid core for the agent loop and an API surface for everything else. By saying no to a hundred things, it delivers something that everyone finds usable. For everything else, "There's an extension for that". That gives you a sense of ownership and doesn't lock you into bad decisions.
Tau builds on that idea and attempts to do the same at the UI level. Instead of shipping every feature, it focuses on the core things — sessions, models, and transcripts — and does them well. Something that you use 10 hours a day has to be fast, reliable, and provide a feeling of ownership.
Coding is changing every month now. The best coding tools are simple at the core and extensible at the edges. A lot of agent orchestration tools make a mistake of rushing to ship as many features with LLMs as physically possible. Half of those features become redundant within a month. The work is getting increasingly high level every month. Things like sophisticated code review tools, subagent visibility, and built-in VCS and project management integrations make less sense as models become more capable.
Also, no tabs.