A browser agent that brings its own hands, brain, and eyes

What it does

You give Scout a task in natural language — “find the top three stories on Hacker News” — and it opens a real browser, reads pages, clicks, types, and scrolls until the job is done, then returns the answer. If it can’t finish, it says what’s missing and hands back whatever it found. Answers can be forced into a JSON Schema shape you provide.

The interesting bit

Scout has no built-in model and no default. Instead it splits the work into four jobs — hands (per-step actions, wants a fast model), brain (planning for long multi-page tasks), eyes (vision for pages that must be seen), and extract — and lets you assign a different model, endpoint, and key to each. That’s a refreshingly honest architecture: the agent loop is the product, the intelligence is yours to rent or self-host.

Key highlights

• Runs on Outcrawl’s cloud (residential connections, captcha solving, step-by-step run replays) or on your own machine via Playwright or any Chrome over CDP.
• Every run returns a structured result: status, data, error, per-step timings, costUsd, and a replayUrl on cloud runs.
• Caps on steps, duration, and budget end runs gracefully with partial results rather than crashing.
• The repo is a monorepo: the agent loop (runtime), model plumbing (model), and CDP browser driver (scout-cdp) are separate packages you can read individually.
• Apache-2.0, with agenthands (input planning) under MIT.

Caveats

• The 92.5% benchmark figure is Outcrawl measuring its own product on Browser Use’s benchmark — self-reported, though they publish the method and every run.
• This GitHub repo is an automatic mirror of the production repo; PRs land via periodic sync commits, so the loop is slower than a typical open-source project.
• Cloud features (replays, captcha solving) are tied to Outcrawl’s paid service; local runs leave replayUrl empty.

Verdict

Worth a look if you want a browser agent whose model layer you fully control, or a readable reference implementation of an agent loop. If you wanted a batteries-included product with a bundled model, this is deliberately not that.