Browse.sh releases open library for AI agents navigating websites
Original: I’ve been juicing up my openclaw w so many http://browse.sh skills - I’ve been delegating all micro life decisions to my claw and browse!
Deep summary
What's new: browse.sh is an open-source CLI tool that wraps browser automation and a curated catalog of site-specific "skills" into a single interface designed to be driven by AI agents. Skills are pre-built adapters for specific websites — Airbnb, Amazon, AllTrails, Craigslist, and roughly 40+ others at launch — that surface structured JSON outputs and encode the correct DOM selectors or XHR endpoints so an agent does not need to discover them by trial and error.
How it works: Each skill is classified by its access strategy: url-param (query-string manipulation), api (public REST endpoints), browser (full headless Chromium via Playwright-style automation), or hybrid (combination). The CLI exposes low-level primitives — click, scroll, type, hover, press — with elements addressable by or by accessibility references that the agent already holds. Network and console output can be tailed in real time, giving both the agent and the developer a ground-truth view of what the page is actually doing.
The numbers: The headline efficiency claim is a 50x reduction in token consumption relative to naive approaches, achieved by pre-computing DOM selectors and relevant XHR endpoints per skill so the agent never has to receive and parse raw page HTML to figure out how to interact with a site. No independent benchmark is provided; the figure is vendor-stated.
How it deploys: Local sessions run against a bundled Chromium instance. Remote sessions are delegated to Browserbase, a cloud browser infrastructure platform, by prefixing any command with "cloud". Cloud commands include session creation, full-web search, and direct page fetch, with sub-second latency reported (14 results in 1.4 seconds in the demo output shown).
Caveats: The 50x token-cost reduction is unverified and likely depends heavily on task type and baseline comparison; sites with aggressive require stealth sessions, which the tool routes through Browserbase rather than local Chromium, creating a dependency on a paid external service. The skill catalog at launch is consumer-web-heavy and shallow in enterprise or authenticated-workflow coverage. There is no information about how skills handle site redesigns or API changes over time.