Browserbase launched Browse.sh on May 18, an open catalog of more than 100 curated browser skills that any AI agent can install with a single CLI command. The catalog solves a problem every production agent team has felt: the discovery tax.

A generic agent loop searching Craigslist listings costs roughly $0.22 per run, according to Browserbase’s benchmarks. The agent has to discover that the search page is fully JS-rendered, stumble onto the hidden JSON API at sapi.craigslist.org, figure out the positional array decoding, learn that item[0] is an offset (not the posting ID), and work around IP-based geo-scoping. Every run pays that discovery tax from scratch.

After four Autobrowse iterations, the graduated Browse.sh skill does the same job for $0.12 per run. That is a 45% cost reduction from better memory.

The unit economics are brutal for anyone running browser agents at scale. Reasoning has stopped being the constraint. Memory has become the bottleneck. Browse.sh is Browserbase’s bet that the industry needs a shared place to store what agents learn about real websites.

What a skill actually is

A skill is a markdown file (SKILL.md) plus any helper scripts needed to repeat a browser workflow reliably. It contains the exact steps, gotchas, API endpoints, selectors, and fallback strategies an agent needs to complete a task on a specific site. No vector embeddings or screenshot reels. Just plain text that humans can read and agents can execute.

The Craigslist skill, for example, includes site-specific gotchas that took a human engineer hours to reverse-engineer and an agent dozens of dollars in tokens to discover from scratch:

  • Snapshot returns 0 refs on /search/ because the search page is fully JS-rendered. Do not use browse snapshot.
  • item[0] is NOT the postingId. It is an offset from data.decode.minPostingId. Treating it as the ID produces 404s.
  • API geolocates by request IP. Add postal=<zip> to override. A residential proxy is not required.
  • Rate-limit: keep at or below 1 req/s sustained.

Once that knowledge is in a skill, it is free forever.

The Browse CLI (npm i -g browse) is the open-source command-line tool agents use to actually drive browsers, fetch pages, search the web, and load skills on demand. All CLI commands work natively with local Chromium. Users can switch to remote sessions on Browserbase’s Platform by prefixing any command with cloud.

How skills get made

Every skill graduates from Autobrowse, Browserbase’s system that uses AI to improve AI. You give an agent a real task on a real site. It runs the task end to end, studies its own trace, iterates on its strategy, and keeps going until the workflow becomes reliable rather than lucky. Once it converges, it writes out a durable skill.

Users can also generate their own skills by typing any domain and task into browse.sh. Autobrowse runs the task against the live site, iterates until it converges, and publishes the result to the public catalog for anyone to use. This creates a compounding effect: every new skill makes the catalog more valuable, which brings more users, who generate more skills.

The catalog at launch spans marketplaces (Craigslist, Zillow, Amazon, eBay), food and dining (OpenTable, DoorDash, McDonald’s online ordering), travel (flight search, hotel booking, Airbnb), government (federal grants portals, state program catalogs), developer tools (GitHub, npm, documentation sites), and enterprise SaaS via partner integrations. Partner skills from companies like Ramp, Lovable, Poke, and Reducto ship with a verified badge.

The industry is converging on this

Claude Code ships with skills. OpenAI Codex supports them. The AgentSkills standard is gaining traction. Every major agent framework is adding some version of “load a markdown file that tells the agent how to do a specific thing.”

Browser skills are the natural next step. The web is messy: sites render differently for different user agents, gate content behind JavaScript, hide data behind undocumented endpoints, throw CAPTCHAs on a whim, and redesign their flows on a Tuesday. A generic agent loop copes with all of that in the moment, then forgets everything once the session closes.

Browse.sh captures what the agent learned, so the next agent (or the next teammate, or the next customer) does not have to learn it again.

The dominant story about browser agents right now is that they will get good when the underlying models get good. Browserbase does not entirely buy that. Even a perfect model still has to discover, on every new site, what a perfect model would already know if it had been there before. Without a place to put what the agent learns, every run is a fresh start.

What this means for AI builders

For teams running browser agents in production, the cost math is the headline. A 45% reduction per run on a task that runs thousands of times a day changes the economics of what is worth automating. The difference between $0.22 and $0.12 per run is the difference between a cost curve that flatlines and one that compounds against you.

But the deeper implication is about memory architecture. The industry has spent the last two years optimizing reasoning: better models, better prompting, better tool-use. Browse.sh suggests that the next frontier is memory, in a form that humans can audit and agents can execute. A skill is not a vector database. It is not a fine-tuned model. It is a markdown file that captures the shortest reliable path through a real website.

That is a bet on plain text as the interface between agents and the web. It is also a bet that the web is messy enough that even the best model needs a playbook.

The bottleneck for browser agents was never intelligence. It was amnesia. Browse.sh is the cure, but only if the catalog keeps growing faster than the web changes.