Set up your AI provider

Every AI feature — commit messages, explanations, reviews, the daily briefing, commit search — talks to one provider you configure once. The app ships no cloud AI and no key of its own: your prompts (which include diffs and commit messages from the repository you point it at) go to the endpoint you chose, and nowhere else. See Private by design.

Option 1 — Ollama (local, the default)

The app is preconfigured for Ollama at http://localhost:11434:

brew install ollama
ollama pull qwen2.5-coder

Then open Settings → AI (⌘,), pick the model you pulled in the model dropdown, and press Test connection. The footer pill turns green ("Connected") when the provider answers — that's it, every AI feature is live, fully offline.

Any model works; coder-tuned models give noticeably better commit messages and reviews. If your machine can hold two, the optional fast model field lets the app use a smaller one for the high-volume map phases (per-file summaries) and the main one for final compositions.

Option 2 — any OpenAI-compatible endpoint

Choose the OpenAI-compatible preset in Settings → AI and fill in the base URL, the model name, and an API key if the endpoint wants one. This covers LM Studio, llama.cpp's server, vLLM, a team gateway, or a commercial API — anything speaking /v1/chat/completions.

The key is saved to your macOS Keychain rather than to a settings file, so the field shows "a key is saved in your keychain" instead of the key itself — there is nothing to read back, which is the point. Type a new one to replace it, or use Remove to delete it. See Private by design.

Remote endpoints see your diffs

With a non-local URL, AI requests leave your machine. The rest of the app stays fully local either way — only the AI features use this endpoint.

If something doesn't work

• "AI provider is unreachable" banner — the provider isn't answering GET /v1/models: check it is running (ollama list), and that the URL in Settings matches its port.
• Generations cut off or oddly short — the model's context window is too small for the diff; the app trims what it sends, but a bigger-context model helps for large changes.
• Everything is slow — local generation speed is your hardware's; the fast-model field and smaller models are the lever.