browse.sh is two things: an "open web catalog" of reusable browser skills, and the browse CLI that consumes them. Both come from Browserbase. The pitch is one command-line tool that AI agents (Claude Code, Cursor, etc.) call to fetch pages, search the web, drive a real browser, and load site-specific automation recipes on demand. The name is misleading: this is not the old terminal browser Browsh (brow.sh). Different project entirely.
The problem it solves
Browser agents redo the same discovery on every run. Each task re-learns the page structure, the selectors, the hidden API calls, burning tokens to rediscover what was already figured out last time. Browserbase benchmarked a Craigslist search at ~$0.22 per run without a skill versus ~$0.12 with one. The catalog claims suggested DOM selectors and XHR requests cut token cost by up to 50x on some tasks.
What a skill is
A skill is a plain-text playbook: a SKILL.md file plus any helper scripts. It encodes the exact steps, gotchas, API endpoints, selectors, and fallback strategies an agent needs to complete one task on one site. Human-readable, not embeddings or screenshots. The skills are generated by Autobrowse, Browserbase's system that iterates on a real task until it converges on the cheapest, fastest path. The catalog launched with 100+ curated skills spanning marketplaces (Craigslist, Zillow, Amazon), travel (OpenTable, AllTrails), government portals (recreation.gov, weather.gov), and SaaS (Ramp). Missing site? You can request generation.
Install a skill, then let the agent use it:
browse skills add alltrails.com
browse skills add recreation.gov
browse skills list
claude "Plan a road trip to Utah with EV charging stops and campsites. Book and reimburse on Ramp."
The CLI
npm i -g browse
Package is published as browse (v0.8.x as of writing), described as the "Unified Browserbase CLI for browser automation and cloud APIs." Command surface splits into a few groups.
Browser primitives drive any page by CSS selector or by the agent's accessibility refs (@8 style):
browse click "input#search"
browse type "Apartments in SF"
browse select @8 "Single family unit"
browse press "Enter"
browse mouse scroll 50 50 10 10
browse screenshot
Debugging tails network and console in real time, so the agent sees exactly what the page did:
browse network --tail
browse console --tail
GET /api/listings 200 124 ms
POST /api/filter 200 84 ms
error Uncaught TypeError: cannot read 'id'
Other groups: snapshot / refs / highlight (inspection), fill / upload (input), eval / get / is (utilities), plus tab, templates, and functions (deployable serverless automations).
Local vs cloud
All commands run against local Chromium by default. Prefix any command with cloud to run it on a remote Browserbase session instead. Cloud also exposes Browserbase's Search and Fetch APIs.
browse cloud sessions create
browse cloud search "Latest White House press release"
browse cloud fetch "https://www.nytimes.com/section/us"
Cloud requires a key: export BROWSERBASE_API_KEY="...". Persistent state (cookies, localStorage) carries across runs via contexts, and fetch falls back from lightweight HTTP to full browser rendering when a page needs JS. Under the hood it leans on Stagehand, Browserbase's AI browser-control SDK.
Pricing
The CLI and the catalog are free. Browserbase markets the browse CLI as open-source, but the repo the npm package points to (github.com/browserbase/cli) returns 404, so the source is not actually public as of May 2026. Local use needs nothing beyond Node and Chromium. Cloud sessions bill against a Browserbase account (Developer plan ~$20/mo, 25 concurrent browsers, 100 browser hours, $0.12/extra hour).
Limitations
- "Open-source" claim does not hold up: the linked CLI repo is 404, no public license.
- Anything useful beyond local browsing pushes you onto Browserbase cloud, which is paid and metered by browser-hour.
- Skill quality depends on Autobrowse and on the site not changing. A site redesign breaks the recorded selectors until the skill is regenerated.
- Single vendor. The catalog, the CLI, the cloud, and the underlying SDK are all Browserbase, so this is a funnel into their platform.
How it relates to my other notes
Complements !ai/browserbase-skills (the browserbase/skills repo bundling Claude Agent SDK) and !apps/browserbase (the cloud browser infrastructure). browse.sh is the agent-facing front door: the catalog plus the browse command that ties the skills and the cloud together.