Search

A protocol-first AI coding harness for running, supervising, and reviewing agent work in isolated git worktrees.

Pronounced "troov" (rhymes with groove; French /tʁuv/). Trouver is French for "to find." The full story is in NAME.md.

Looking for the fast, incremental code-search tool? See trouve-search and its agent installation guide.

What trouve provides

• Isolated sessions. Each unit of work owns a git worktree and branch; parallel threads can collaborate inside the session.
• A choice of agents and models. Run Codex, Claude Code, or Cursor through their supported CLIs and subscription authentication, use hosted API providers, or run supported local models.
• Visible, reviewable work. Stream chat and tool activity, handle approvals, inspect diffs and files, use multiple integrated terminals, and create or merge pull requests without leaving the session.
• Extensible workflows. Add MCP servers and skills, schedule automations, and run self-hosted GitHub App-backed pull-request reviews.
• One responsive product UI. The Lit application powers both the Wry desktop client and the mobile-oriented PWA.

The product is chat-first rather than IDE-first: files and terminals support the agent workflow, while the diff remains the primary review surface. See the UX screen map for the complete interaction model.

Architecture

Clients communicate with trouve-server exclusively through versioned HTTP and SSE APIs. Durable user-visible state is reconstructed from one persisted, cursor-addressed event log, and every agent side effect passes through the same permission and audit boundary. The desktop application embeds a local server but still talks to it over the protocol.

Area	Location
Desktop product host	crates/trouve-app
Shared Lit desktop/PWA frontend	web/app-ui
Protocol server and harness engine	crates/trouve-server, crates/trouve-core
Provider and vendor-agent integrations	crates/trouve-providers, crates/trouve-agents
Published semantic code search	crates/trouve-search

The repository invariants and a crate overview are in AGENTS.md; architectural decisions are recorded in docs/adr.

Run from source

The workspace currently requires Rust 1.92 and Node.js 24, plus the native system dependencies required by Wry on your platform.

Start the frontend development server:

npm --prefix web/app-ui ci
npm --prefix web/app-ui run dev

Then launch the desktop host from another shell:

TROUVE_APP_UI_DEV_URL=http://127.0.0.1:5173 cargo run -p trouve-app

To build a release binary with the desktop assets embedded:

npm --prefix web/app-ui ci
npm --prefix web/app-ui run build
TROUVE_APP_UI_DIST="$PWD/web/app-ui/dist/desktop" \
  cargo build --release -p trouve-app

Development

The main workspace checks are:

cargo fmt --all --check
cargo clippy --all-targets -- -D warnings
cargo test --workspace
npm --prefix web/app-ui run format:check
npm --prefix web/app-ui run lint
npm --prefix web/app-ui test
npm --prefix web/app-ui run build

Model-downloading and network tests stay ignored unless their documented environment flags are enabled. Search-specific parity and benchmark commands live in the trouve-search README.

Useful design and operations documentation:

• Event-log design
• Provider architecture
• MCP, skills, and GitHub integrations
• Self-hosted pull-request reviews
• Changelog

License

MIT. The trouve-search package contains portions derived from MinishLab/semble; see its acknowledgements.