Outcrawl released Scout, an open-source browser agent, with a benchmark

A browser agent is still the practical answer when an AI product has to work on a site that offers no API: it loads the page, reads it, clicks, types and reports back. Outcrawl, which has sold cloud infrastructure for this work since earlier this year, released Scout on 13 October as an open-source agent that runs the same loop its paid service does, under Apache 2.0. The repository passed 740 stars this week.

The four roles the agent uses

Scout splits the work across up to four model roles, and only the first is required. Hands reads the page and takes each step: click, type, scroll, read. Brain, optional, plans the run and decides each step. Eyes, optional, looks at a screenshot when the page cannot be read without one. Extract, optional, reads the page into an answer or a schema the caller gives. Each role is any model that speaks the standard chat completions API, so the user brings their own keys from whichever provider.

Where the agent runs

Three ways. On Outcrawl's cloud, which is Outcrawl's commercial product and gives every account 1,000 free credits a month without a card. On a local Playwright process, which Scout launches for you. Or on an existing Chrome or Chromium through CDP, which lets Scout drive a browser the user is already signed into. Every run is recorded by default, and result.replayUrl opens a step-by-step replay of what Scout did.

The benchmark, which is Outcrawl's own

The README reports that on three full runs of Browser Use's 80-site benchmark, dated 7 October 2026, Scout running on Outcrawl's cloud reached 92.5 percent of the pages and "the next best" reached 81.4 percent. The benchmark harness belongs to Browser Use, a competitor; Outcrawl ran its own agent on it. The second score is unlabelled in the README, so a reader cannot tell which service it refers to from this source alone. Outcrawl's site names the comparison more precisely, but the benchmark remains the vendor running its own measurement, and the number should be read that way.

What is actually open source

Scout the agent, under Apache 2.0, with packages/agenthands kept under MIT, copyright Aeonmind LLC. The hosted cloud, the residential-connection network, the captcha-solving step and the session replay viewer are Outcrawl's commercial product and are not in the repository. A team that wants to self-host gets the planner and the browser loop, not the surrounding reliability work that Outcrawl sells.

The usefulness of this release is less in the benchmark number and more in the choice. A team that is picking between browser-agent libraries now has a current, maintained, commercially backed open-source option alongside Browser Use itself and the smaller harnesses. Starting with the open repository, running the Playwright mode on your own machine, and moving to the paid cloud only if the reliability work matters is a reasonable path, and the shape of the repository supports it.

Source

• Outcrawl, Scout
• Outcrawl