outcrawlai/scout

Scout is worth a look if you want a browser agent you can embed in TypeScript code without committing to one model vendor. Expect to supply both a model and a browser, and to treat anti-bot handling as an operational problem rather than a solved one.

The problem

You have a web task that only exists as a page: pull the top three Hacker News threads with their comment counts, or place an order through a checkout form. Today you either write brittle selectors against a page that keeps changing, or you hand the job to a general chat assistant that cannot act. Scout changes the step where the actions happen: you describe the task in words and pass a schema, and the agent reads the page, clicks, types and scrolls through a real browser until it returns data in that shape. When it cannot finish, you get a partial result plus an error stating what is missing.

What's different

The design splits model work into separate jobs rather than using one model for everything. A required hands model reads the page and decides each click, type, scroll or read on every step; optional brain and eyes models handle multi-page planning and screenshots. You choose a model per job on any service that speaks the standard chat completions API, including something running locally, and there is deliberately no built-in default. The practical benefit is cost control: you can put a fast, cheap model on the high-frequency hands loop and reserve a stronger one for planning. The trade-off is that you now own model selection and prompt-tuning quality yourself, and a poor hands choice will show up as bad actions on every step.

Runs are bounded objects. Every call returns a status (completed, partial, capped or failed), data, an error field, a step log with timing, costUsd where known, and a replay link on the cloud. Caps on steps, duration and budget mean hitting a limit is a normal ending with whatever was found so far, which suits scraping-style tasks better than an all-or-nothing script. Browsers are also swappable, from Outcrawl's cloud to a locally launched Chromium to an existing CDP endpoint, which is the same line of config.

Who should try it

Good fit: a TypeScript developer automating a research or form-filling task where the destination page is the only interface, and who wants the agent as a library rather than a chat client. The example scripts in the repository give runnable starting points. Poor fit: anyone wanting an unattended production pipeline with no model configuration, or someone whose tasks run on sites that block automation, since the README points to Outcrawl's cloud for that. If you already have your own Chrome and a model endpoint, the self-hosted path is viable; check the setup instructions and the package reference first.

Requirements and limitations

Node 22 or later is required. The project has no model of its own and no default, so you must supply at least an hands model endpoint and key before anything runs; the optional jobs each need compatible models, and eyes requires image input. If you use Outcrawl's cloud you need an Outcrawl key, which the README says comes with a free monthly credit allowance; self-hosting avoids that key but puts browser anti-bot handling on you, and the README explicitly frames many sites as hostile to data-center or script-like browsers. Ongoing upkeep is mostly model configuration and the quality of the prompts for the job split.

The repository is organized as several packages, and packages/agenthands carries a different MIT licence from the Apache-2.0 main licence, which matters if you vendor components. The README also says the GitHub copy is mirrored automatically from Outcrawl's own repository, so contribution and issue flow may not follow normal GitHub patterns.

HotRepo's take

The repository was created 2026-10-13 and appeared on HotRepo's NEW-REPOSITORY board for TypeScript at #9 with 937 total stars, observed 2026-10-25. That is a new-project board rank, not an overall Trending figure, and HotRepo has no comparable daily history, so persistence cannot be judged from these samples. The more useful pre-adoption check is whether the hands model you intend to use is fast enough and whether your target sites work without the cloud browser, since that combination determines the real cost of a run.