agent-model

# Choosing the browser-use agent brain The `retry_with_browser_use_agent` tool and cloud sessions run an **autonomous** agent loop that needs its own LLM to think with (it makes its own API calls — it does not borrow this Claude session). This skill decides *which* LLM, per task. The Python MCP server is deliberately dumb: it reads config and accepts an override. The judgment lives here. ## Selection precedence (how the server resolves the brain) 1. **Session override** — set at runtime by the `browser_set_agent_model` tool (this turn only). 2. **Configured** — the `"browser-use"."agentModel"` block merged from settings files: `~/.claude/settings.json` → `<project>/.claude/settings.json` → `<project>/.claude/settings.local.json` (local wins). 3. **Legacy** — if `BROWSER_USE_API_KEY` is set and nothing above, Browser Use's `bu-latest`. 4. **Default** — `anthropic` / `LATEST_SONNET_MODEL` (uses `ANTHROPIC_API_KEY`, already available). ## Providers | provider id | use for | key (env var NAME) | base_url | |---|---|---|---| | `anthropic` | Claude models — strong vision, best default | `ANTHROPIC_API_KEY` | not needed | | `openai` | GPT models | `OPENAI_API_KEY` | not needed | | `openai_compatible` | Kimi, GLM, Moonshot, Zhipu, or any OpenAI-compatible endpoint | your choice | **required** | | `browser_use` | Browser Use's hosted BU3 models (`bu-latest`, `bu-2-0`) | `BROWSER_USE_API_KEY` | not needed | Current Anthropic vision models: `LATEST_SONNET_MODEL` (default; best multi-step + vision), `LATEST_OPUS_MODEL` (hardest reasoning), `LATEST_HAIKU_MODEL` (cheapest/fastest). Confirm live ids against the `claude-api` skill before quoting one — do not rely on memory. ## Picking a model for the task - **Simple scrape / read one page** → cheap+fast is fine: `LATEST_HAIKU_MODEL`, or a cheap `openai_compatible` model. - **Multi-step form fill, navigation, recovering from popups** → `LATEST_SONNET_MODEL` (default). - **Login / auth flows, ambiguous pages, hard reasoning** → `LATEST_SONNET_MODEL` or `LATEST_OPUS_MODEL`. - **CAPTCHA / stealth / anti-bot** → a cloud session (`browser_start_cloud_session`) with `browser_use` / `bu-latest`, which is tuned for it. - If the user names a model or provider (Kimi, GLM, "the cheap one", "smartest"), honor it. **Suggest the right model even when it differs from their configured default** — that's the point of this skill. Read the task, recommend, then apply one of the two ways below. ## Two ways to apply a choice **Ephemeral (this session, reverts next launch)** — call the tool: ``` browser_set_agent_model(provider="anthropic", model="LATEST_SONNET_MODEL") browser_set_agent_model(provider="openai_compatible", model="moonshot-v1-128k", base_url="https://api.moonshot.cn/v1", api_key_env="MOONSHOT_API_KEY") ``` No file edit; takes effect on the next agent call. **Persistent (per project or per user)** — write to settings (relaunch the MCP server after). Use the `update-config` skill to edit `.claude/settings.json` (committed) or `.claude/settings.local.json` (gitignored): ```jsonc { "browser-use": { "agentModel": { "provider": "openai_compatible", "model": "glm-4.6", "baseUrl": "https://open.bigmodel.cn/api/paas/v4", "apiKeyEnv": "ZHIPU_API_KEY", "temperature": 0.2 } } } ``` ## Secrets — never store keys in settings `apiKeyEnv` holds the **name** of an environment variable, never the key itself (settings.json is often committed). Export that variable in the shell that starts Claude Code. The MCP server inherits Claude Code's environment, so every exported key reaches it, custom providers included. Never add the key to the plugin's `.mcp.json`: a `${VAR}` entry there makes Claude Code refuse to start the server whenever the variable is unset, and edits to an installed plugin do not survive its next update.