autonomous-browser

Plan multi-step browser work using browser-use with an approved model budget.

SKILL.md

GitHub

Autonomous Browser

Scope: browser-use Python library. Source: https://github.com/browser-use/browser-use

Workflow

Use available CLI tools or documented APIs for technical tasks such as deployments, migrations, cloud administration, configuration, builds, logs and database work. A web dashboard alone does not justify a browser workflow. Honor explicit user tool preferences; use this skill only when actual webpage interaction is required.

Use only when deterministic DOM actions fail or multi-step planning is essential. browser-use is a separate Python reasoning engine and may use chargeable LLM APIs. Do not create paid cloud sessions by default. Require an explicit device/browser permission, named objective, URL allowlist, max iterations, timeout, model budget and human checkpoints. Prefer local browser and a caller-provided model account; never assume API credits are free. Inspect planned actions, execute only authorized ones, and save a scrubbed action/evidence log. Stop on CAPTCHA, account lock, MFA, payments, new unknown domains and repeated failures. A user can instead choose the zero-model-cost Browser Operator. No autostart and no implicit installation.

Device and permission gates

Required authorization scopes: browser:read, browser:interact. Scripts, OS/desktop actions, external service side effects, media generation and paid API usage require independent user-approved execution permissions. This skill does not elevate Remote Arc Tool privileges or authorize use of hidden credentials.

Preflight

Check the engines this skill names before relying on them, with tools you already have (for example <engine> --version through start_process, or list_directory on its install folder). Missing executors must be installed explicitly on the correct user device. A Skill being built-in is NOT evidence an engine is present.

Evidence of completion

Report input constraints, executed tools and device, verified output path/result, permissions used, cost (if known) and failures. Do not imply successful execution from merely writing a plan.