Configuration
Open extension Settings (gear icon in the header).
LLM
| Field | Description |
|---|---|
| Provider | Anthropic / OpenAI (or OpenAI-compatible: LiteLLM / Azure / Ollama, etc.) |
| Endpoint | Leave empty for default, or set a custom base URL (e.g. https://api.deepseek.com/v1) |
| Model | Pick from suggestions or type your own (claude-sonnet-4-6, gpt-4o-mini, ...) |
| API Key | Check "session only" to clear on browser close; otherwise stored in chrome.storage.local |
| max_tokens | Per-response cap (default 4096) |
| Max rounds | Max LLM rounds per session (default 20) |
| Optimizer model | Model used by the "optimize prompt" button; empty = use chat model |
| Continuation nudges | If the model stops without calling a tool, ask once more if it's really done (default 1) |
The API Key does not enter IndexedDB and is not included in tool bundle exports.
Appearance
- Theme: dark / light / follow system
- Default view:
- Compact (recommended) — one-line progress per tool call; click a row to expand
- Full — each step shows full args / output
An eye icon in the header toggles per-session without writing back to default.
Permission mode
Toggle in the top toolbar:
- read — only safe tools auto-run
- default — safe + caution auto-run; dangerous requires approval
- trust — safe + caution + allowlisted dangerous auto-run
- yolo — everything auto-runs (careful)
Dangerous tool allowlist
In trust mode you can pick which dangerous tools (like httpRequest(withCredentials)) skip approval.
Coordinator (optional)
Remote WS server URL. When connected the extension accepts remote tool step dispatch. See Coordinator.