Coder

Flutter GUI and a pure-Dart AI coding-agent daemon. The daemon owns workspace, agent, timeline, and approval state; desktop and mobile applications connect to the same versioned WebSocket protocol.

Development

dart pub get --enforce-lockfile
dart run melos verify
dart run melos verify:debug

See docs/testing.md for the individual TDD, contract, widget, golden, coverage, architecture, and Debug E2E commands.

Run the standalone daemon. The seeded OpenAI provider reads OPENAI_API_KEY; additional providers can use an environment variable or a locally stored credential configured from the Settings screen:

dart run melos run run:daemon

The default endpoint is ws://127.0.0.1:7337/ws. Override the state/config directory and listener with TINYRACK_CODER_HOME and TINYRACK_CODER_LISTEN. Without an override, Linux uses the XDG config/state directories, macOS uses Application Support, and Windows uses AppData.

Use the gear button in the connected app to configure multiple OpenAI or OpenAI-compatible endpoints. Each provider can use Responses or Chat Completions, discover model IDs through /models, and retain manually entered models and capability overrides. Explicit model diagnostics make a small streaming tool-call request and may incur provider charges.

Desktop and mobile use separate targets so the mobile bootstrap never starts a daemon. Run these commands from apps/coder_app:

flutter run -d linux -t lib/main_desktop.dart
flutter run -t lib/main_mobile.dart

LAN mode is intentionally plain ws:// for trusted local networks only. Bind a standalone daemon to 0.0.0.0:7337 explicitly and provide a 256-bit token via TINYRACK_CODER_TOKEN. Do not expose this listener to the public internet. Remote clients can inspect the provider catalog but provider and credential mutations are restricted to loopback connections.