browser_session

01 前言 Front Matter

args

[ "{{task}}" ]

name

browser_session

schema

{ "task": { "type": "string", "required": true, "description": "High-level natural language task description. browser-use's\ninternal LLM agent will plan + execute the steps. Examples:\n\n\"Open Hacker News, find the top story right now, return its\n title + URL + first paragraph of the article\"\n\n\"Search GitHub for repositories matching 'computer use agent',\n sorted by stars, return the top 5 with their star counts\"\n\n\"Go to weather.com for 94025, extract today's high/low and the\n 7-day forecast as a markdown table\"\n\nThe task should be self-contained — browser-use can't ask\nfollowup questions, it just plans and acts.\n" } }

command

[ "sh", "-c", "TASK=\"$1\"\n[ -z \"$TASK\" ] && { echo \"task required\" >&2; exit 1; }\n\n# cwd is set to the skill dir by the kernel.\nif [ ! -x ./.venv/bin/python ]; then\n echo \"browser_session: venv missing — first-time setup required:\" >&2\n echo \" cd skills/browser_session && ./setup.sh\" >&2\n exit 1\nfi\n\nif ! ./.venv/bin/python -c \"import browser_use\" 2>/dev/null; then\n echo \"browser_session: venv exists but browser-use not installed —\" >&2\n echo \" re-run: cd skills/browser_session && ./setup.sh\" >&2\n exit 1\nfi\n\n# Sensible defaults so the skill works out of the box on a typical\n# KinClaw setup (local Ollama running, no external API keys\n# required). User-set env (ANTHROPIC_API_KEY etc.) still takes\n# precedence — runner.py picks Anthropic > OpenAI > Ollama.\n: \"${OLLAMA_BASE_URL:=http://localhost:11434}\"\n: \"${BROWSER_USE_MODEL:=kimi-k2.6:cloud}\"\nexport OLLAMA_BASE_URL BROWSER_USE_MODEL\n\n./.venv/bin/python ./runner.py \"$TASK\"\n", "_" ]

timeout

600

description

Multi-step browser automation via browser-use — the open-source framework for autonomous browser navigation (91K+ stars, MIT). Use INSTEAD of `web` when the task spans more than a single page fetch: - Login + navigate + extract (auth state preserved across steps) - 5+ page interaction sequences - DOM-numbered element targeting (the framework shows the agent which element index to click — far more reliable than guessing CSS selectors) - JS-heavy SPAs that need real interaction state, not just rendered HTML Cost: cold start ~10-20s (browser warmup + LLM init); per-step ~2-5s. **Don't use for one-shot fetches** — `web` is faster and cheaper for those. Use this when the task description naturally contains "log in", "navigate to", "fill the form", "find on page X then click", or any sentence with multiple verbs of interaction. First-time setup (one machine, ~5 min): cd skills/browser_session ./setup.sh This creates a per-skill venv at ./.venv/ and installs browser-use + Chromium. Subsequent calls use the venv directly — zero pollution of the system Python. LLM selection (env-driven, in order of preference): ANTHROPIC_API_KEY → Claude OPENAI_API_KEY → GPT-4o OLLAMA_BASE_URL=URL → local Ollama via OpenAI-compat Optional: BROWSER_USE_MODEL=<name> to override the model. Optional: BROWSER_USE_HEADLESS=false to watch the browser run.

02 资源文件 5

• 参考文档references/core/guide.md2.5 KB
• 脚本runner.py6.1 KB
• 脚本scripts/expected_value.py3.3 KB
• 脚本setup.sh2.1 KB
• 参考文档workflows/tuning-pass.md887 B

03 SKILL.md 正文