Browse.sh
Browse.sh is a browser command-line tool intended for AI agents. It combines an open catalog of website-specific skills with commands for interacting with a page, inspecting errors and using remote browser sessions. The publisher shows installation through the npm package named browse.
A developer can use the catalog to give an agent instructions for a particular website instead of asking it to rediscover the same interface during each task. The catalog distinguishes browser, fetch and API methods, with hybrid entries where a workflow combines approaches. These labels describe the published skill; they do not guarantee that an external website will remain accessible.
Start with a narrowly defined website task
The site demonstrates adding skills for individual domains, then passing a task to an agent. Example catalog entries include finding flight status, extracting event listings and searching public contract opportunities. Some entries identify a read-only boundary, while others involve authenticated actions or stop before payment.
Read the chosen skill before running it. A search that returns public information needs different permissions from a workflow that adds an item to a cart or submits a change. Decide which actions the agent may take, and keep account access limited to that task. A skill's presence in the catalog is not permission to act on a user's behalf.
The publisher advertises large token savings from suggested selectors and XHR requests. Those savings are a vendor claim, not a measured result for every website or agent. Evaluate whether a skill reduces repeated navigation in your actual workflow, including recovery when the site's interface changes.
Inspect what the browser did
Low-level commands cover clicking, typing, scrolling, hovering and key presses. Elements can be addressed through selectors or an agent's accessibility references. The screenshot command provides a rendered result to inspect rather than relying only on the agent's narration.
Browse.sh also exposes network and console tails for a session. Those can help distinguish a selector failure from an application error or unsuccessful request. For a task that changes data, verify the resulting page state and the relevant response rather than accepting a successful click as proof of completion.
Separate local browsing from cloud access
The publisher says the commands work with local Chromium. Prefixing commands with cloud switches to Browserbase's platform, including remote session creation and Search or Fetch APIs. A remote session introduces a separate service and access requirements; installing the CLI alone does not establish that cloud usage is free or already configured.
The catalog itself documents login gates and blocked workflows on some sites. Check those limits before promising unattended automation. Do not treat a catalog entry as authorization to bypass an access control or ignore a publisher's terms. Begin with a read-only task, confirm the returned information against the site, and then evaluate any permitted write workflow with explicit checkpoints.