What is Browse.sh and why it exists

Browserbase, an American startup building infrastructure for browser-based AI agents, launched the Browse.sh project on May 18, 2026. It's a combination of two things: a public catalog of browser skills (so-called skills) and an open-source CLI tool that you install with a single command npm i -g browse.

A "skill" in this context is a Markdown file (SKILL.md) that contains precise instructions on how to accomplish a specific task on a particular website. It includes selectors, API endpoints, common pitfalls, fallback strategies, and specific quirks of the given page. No vector embeddings, no screenshots — just plain text that both humans and agents can understand.

The project is run by a team consisting of Kyle Jeong, Shubhankar Srivastava, Alex Qiu, and Shrey Pandya. The catalog launches with 100 skills covering marketplaces (Craigslist, Amazon, eBay), travel (Google Flights, Airbnb, Booking.com), dining (OpenTable, DoorDash), government services, and developer tools (GitHub, npm).

The problem Browse.sh solves: agent amnesia

AI agents, whether in Claude Code, Cursor, or Codex, today share one fundamental shortcoming. Every time they get a task on the web, they have to "rediscover" the page. They open the browser, search for the right button, click, parse the response — and once the session ends, they forget everything. The next run starts from scratch again.

Browserbase calls this phenomenon the "discovery tax" — the cost of exploration you pay with every agent run. On a Craigslist benchmark, the team measured that a generic agent pays roughly $0.22 for a single listing search. It has to figure out on its own that the results page is fully rendered with JavaScript, discover a hidden JSON API at sapi.craigslist.org, understand positional field decoding, and solve geolocation restrictions based on IP address.

However, after four iterations of the Autobrowse system (more on that below), the resulting skill performs the same task for $0.12 — that is, 45% cheaper. "Every subsequent run is cheaper because the skill contains the shortest reliable path the agent found and reuses it instead of re-deriving it," explains Kyle Jeong on the Browserbase blog.

How Browse.sh works

The basic workflow is surprisingly simple:

1. CLI installation: npm i -g browse
2. Skill selection: On the browse.sh website, find the skill for your target site (e.g. zillow.com/extract-listings)
3. Skill installation: browse skills add zillow.com/extract-listings
4. Agent execution: The agent loads the SKILL.md, follows the described procedure, and returns a structured result

A typical prompt then looks something like: Use /extract-listings to find apartments under $3,000 in SF with 2+ bedrooms. The agent reads the skill, executes the workflow according to the instructions, handles edge cases based on documented "gotchas," and returns structured data.

Each skill contains sections like Site-Specific Gotchas — for example, the Craigslist skill notes that item[0] is not a posting ID but an offset, and that the API geolocates by IP address (overridable with the postal=ZIP parameter). These are insights that would take a human developer hours of reverse engineering and an agent tens of dollars in tokens.

Autobrowse: AI that improves AI

The key technology behind Browse.sh is the Autobrowse system, which Browserbase introduced in early May 2026. It's a self-improving loop: an agent receives a real task on a specific website, executes it, reads its own trace, and iteratively refines its strategy until it achieves a reliable result.

"Instead of clicking step by step, the agent decides to evaluate JavaScript directly on the page and save this workflow as a skill," Kyle Jeong described on X. In the Hacker News example, two Autobrowse iterations shortened the time from 102 to 35 seconds, the number of steps from 23 to 8, and the cost from $1.46 to $0.28.

The resulting skill is then published to the open catalog at browse.sh, where it's available to anyone. If the desired website isn't in the catalog, simply enter the domain and a task description on browse.sh — Autobrowse will generate and publish the skill.

Who is Browse.sh for

The primary target audience is AI developers building agents for automating web tasks — whether it's data extraction, form filling, testing, or monitoring. Also product teams that deploy browser features and need deterministic, auditable playbooks instead of black-box agent runs.

Browse.sh is open-source and free. The skill source code is available on GitHub, where the browserbase/skills repository has already collected over 3,400 stars. The catalog uses the AgentSkills standard, which is gaining traction across the ecosystem — supported by Claude Code, OpenAI Codex, and other agent frameworks.

Czech and European context

Browse.sh is a globally available tool without language barriers — skills are written in English, but the concept of an open skill catalog itself can be used by any Czech developer working with AI agents. For Czech companies experimenting with Claude Code, Cursor, or Codex (for example, for test automation or web data extraction), Browse.sh represents a path to significantly lower operational costs.

In the European context, it's also important that Browse.sh is open-source and transparent — skills are readable by both humans and machines, which facilitates auditability in compliance with the EU AI Act. Czech startups and development teams building AI solutions for the European market can thus integrate Browse.sh without concerns about regulatory complications.

Browserbase, the parent company behind Browse.sh, is an American startup from San Francisco. In addition to Browse.sh, it operates a cloud platform for browser automation that includes session management, automatic CAPTCHA solving, residential proxies, and other tools for reliable web browsing. Their customers include companies like Ramp, Lovable, Poke, and Reducto.

What this means for the future of AI agents

Browse.sh arrives at a time when agentic AI is going mainstream. As Kyle Jeong astutely notes: "The bottleneck for browser agents was never intelligence. It was amnesia." Even if we had a perfect model that flawlessly understands every web page, it would still have to discover each page anew — unless it had somewhere to store what it learned.

Browse.sh solves this problem elegantly: memory that compounds. Each new skill enriches the catalog, each new user brings more websites, and each subsequent agent run is cheaper than the previous one. At a time when companies count every dollar in tokens (and prices of models like GPT-5.5 or Claude Opus 4.7 are certainly not falling), this is a competitive advantage that cannot be ignored.