A desktop and browser control plane for orchestrating AI coding agents.

Racc lets you plan, launch, monitor, test, and merge work produced by terminal-based AI coding agents. Every normal task can run in its own git worktree and native terminal session, so multiple agents can work in parallel without sharing a working tree.

Racc is a control plane, not a code editor. It currently supports Claude Code and Codex in both local and remote workflows.

• Task Board — Four focused columns: Open, Working, Merge Manager, and Test Manager. Completed tasks are archived instead of occupying a visible Closed column.
• Task Planner — Give Claude Code or Codex an Epic link or product description, review the generated dependency-aware plan, and create only the tasks you select.
• Isolated agent sessions — Fire a task into the repository or a dedicated git worktree, with an interactive xterm.js terminal and persistent session metadata.
• Merge Manager — Queue PRs from Working, configure the target branch and instructions per repository, test the combined tree, and merge the queue in order.
• Test Manager — Run an isolated, read-only full-project UAT pass. The default prompt is editable per repository and the action button is simply Run.
• Structured manager results — Planner, Merge Manager, and Test Manager report through capability-scoped loopback MCP tools. Racc stores the result directly and refreshes the UI; printed JSON is not used as a completion protocol.
• Direct terminal links — HTTP(S) links open in the system browser without xterm's generic warning dialog. Detected file paths open in Racc's file viewer, including optional line numbers.
• Remote servers — Run agents over SSH in persistent tmux sessions and reconnect from the dashboard.
• Headless server — racc-server serves the same React UI and WebSocket transport for browser access, including terminal streaming.
• Usage visibility — Aggregate Claude Code token usage and automatically configure RTK output compression when available.

See the User Guide for the end-to-end workflow and recovery behavior.

Grab the latest macOS .dmg from the Releases page. Linux and Windows users can currently build from source with the Tauri prerequisites for their platform.

Prerequisites: Rust (stable), Bun (v1.0+), Git, and the Tauri 2.x prerequisites for your platform.

git clone https://github.com/liu1700/racc.git
cd racc
bun install

# Desktop app
bun tauri dev

# Headless server (browser access)
bun run build
cd src-tauri
RACC_DIST_PATH=../dist cargo run --bin racc-server
# Open http://localhost:9399

racc-server supports RACC_PORT (default 9399), RACC_DB_PATH (default ~/.racc/racc.db), and RACC_DIST_PATH (default dist). It binds to 0.0.0.0 and currently has no application-level authentication or TLS. Expose it only on a trusted network such as a private Tailscale tailnet.

bun test
bun run build

cd src-tauri
cargo check --workspace
cargo test -p racc-core

Racc uses a three-crate Rust workspace:

• racc-core owns business logic, SQLite, local PTY and SSH/tmux transports, workflow managers, and events.
• racc-server is an Axum HTTP/WebSocket server for the browser build.
• The Tauri desktop crate provides thin command wrappers and native desktop integrations.

The shared React frontend talks through RaccTransport, which selects Tauri IPC in the desktop app or WebSocket in a browser. See Technical Architecture, Session Lifecycle, and WebSocket Remote API.

Short for raccoon — clever, adorable, with nimble little hands. But be careful: they can be surprisingly brutal sometimes.

Contributions are welcome. See CONTRIBUTING.md for setup, architecture, and verification guidance.

MIT