| name | browser |
| description | Drive the user's real browser (Chromium/Brave/Chrome) — open url, web search, read page text, click, type, screenshot, list/switch tabs. Real backend = Playwright connect_over_cdp (attaches to your existing browser); mock for the binding benchmark. Pure execution (slot-filled, no LLM). |
Browser Skill
Deterministic control of the user's own browser. Canonical reflexer case — each
request is one slot-filled command, no autonomous reasoning.
Backends. Default MOCK echoes JSON (binding-only). The REAL backend uses
Playwright connect_over_cdp to attach to a running Chromium/Brave/Chrome started
with --remote-debugging-port=9222 (preserves your logins/tabs). On Raspberry Pi /
ARM, attach to the apt chromium-browser (never Playwright's bundled binary); raw CDP
(pychrome) is the lightweight fallback. See docs/research/browser_skill_options.md.
Commands (via bash)
CLI: python3 agents/skills/browser/browser.py <command> (aliased browser).
browser open --url "<url>"
browser search --query "<text>" [--engine google]
browser read-text [--selector "<css>"]
browser click --target "<css selector or visible text>"
browser type --selector "<css>" --text "<text>"
browser screenshot [--path <file>] [--full-page]
browser tabs
browser tab-switch --id <id>
Output format
One JSON line: {"ok":bool,"data":{...},"error":null|{...},"meta":{...}}. Exit 0/2.
On an ambiguous click target the real backend returns error.code = "ambiguous"
rather than guessing.
Pure-executioner boundary
No smart-waits, no "do this whole task" verbs, no arbitrary JS-eval, no credential
handling — those would require a thinker. The browser is just one more OS surface the
reflexer fills slots against.
Collisions (why this skill is in the benchmark)
browser open --url vs app open --name (URL vs app); browser search vs maps search
(web vs places) vs spotify search (web vs music); browser tab-switch vs app switch.
