Try it
Desktop app: download the .dmg (macOS), .AppImage (Linux) or .exe (Windows) from
Releases, then add a provider and a model in
Settings (Getting started).
A team of agents: From one agent to a team
asks a lead, an analyst and a reviewer why revenue fell in a sales export. You watch the live
agent tree, approve the file a worker writes, read the bill per model and mix a local model
into the team. Agent teams are supported and experimental.
Terminal, in one line: chatty-tui needs no config to talk to a model server you already run:
chatty-tui --ollama # local models via Ollama
chatty-tui --openai-compat-url http://localhost:8000 # vLLM, llama.cpp, LM Studio, …
git diff main | chatty-tui --pipe # stdin in, answer on stdout
chatty-tui ships with the desktop app (Install CLI…) or builds with cargo build --release -p chatty-tui.
In your editor: chatty-tui acp speaks the Agent Client Protocol, so Zed can use Chatty as its agent.
What it does
- Works on your machine. It uses files, a persistent shell, git, SQL over CSV/Parquet/Excel (DuckDB), web search and fetch, and a real Chrome it drives and you can take over. The shell runs in a sandbox on Linux (bubblewrap) and macOS (sandbox-exec), and approval modes gate every side effect.
- Lets you watch. A terminal dock under the chat pins the agent's own shell as a tab you can type into. Edits render as diffs, multi-step work as a live to-do plan, and every reply shows its token cost.
- Produces files. It typesets PDFs with Typst, draws charts and query tables, and reads and writes Word, Excel and PowerPoint, all rendered in a panel beside the chat.
- Checks sub-agents instead of trusting them. In a git repository, each worker runs in its own worktree. Chatty runs your verification command itself and attaches the result to the worker's answer. Roles (
coordinator, coder, reviewer) and --team presets are declared in files you commit.
- Uses what you already have. It reads
AGENTS.md / CLAUDE.md, SKILL.md skills, and MCP servers.
- Stays local-first. Conversations are stored in SQLite and memory in a local file. There's no telemetry: traffic goes only to the providers and services you configure, plus GitHub for update checks. It runs fully offline with Ollama.
How it's built
A Cargo workspace where the desktop app, the terminal app and the editor server all drive one agent core, so they behave the same:
chatty-core: the agent: turn loop on rig, tools, sandbox, memory, settings, persistence.chatty-gpui: the desktop app (chatty) on GPUI, the GPU-accelerated UI framework from Zed.chatty-tui: the terminal app on Ratatui: interactive, --headless, --pipe, and acp.chatty-terminal: the PTY and terminal emulation behind the dock and the agent's shell.- WASM modules: agents as
wasm32-wasip2 components (Wasmtime runtime, module SDK, and an OpenAI/MCP/A2A gateway).
Full map: AGENTS.md · workspace crates · system overview.
Build from source
make setup # Linux system deps + wasm32-wasip2 target (once)
make build
cargo run -p chatty-gpui # desktop app
cargo run -p chatty-tui -- --ollama # terminal app against a local model
make test-fast # chatty-core unit tests
make ci # what CI runs for Rust PRs
A full test build needs about 16 GiB of target/ (why).
Walkthrough: Build and run in 10 minutes.
Contributing and extending
License
MIT
