Unbrowse is a drop-in replacement for browser automation in agent stacks. It learns the request path behind websites and reuses it as a skill, so agents stop repeating the same browser workflow every time.

How does Unbrowse work?

Unbrowse can use a real browser on the first pass to learn the request flow behind a site, then reuse that learned route as a skill on later runs. The browser stays available for auth and hard cases, but repeat work stops depending on the DOM.

Which AI coding agents work with Unbrowse?

Unbrowse works with all major AI coding agents including Claude Code, Cursor, OpenClaw, and Windsurf. The simplest full install path is curl -fsSL https://unbrowse.ai/install.sh | bash. After install, hosts with skills support can also use npx skills add unbrowse-ai/unbrowse.

Is Unbrowse free to use?

Yes, Unbrowse is free and open-source. Install it with curl -fsSL https://unbrowse.ai/install.sh | bash and start discovering APIs immediately. After install, hosts with skills support can also use npx skills add unbrowse-ai/unbrowse. The shared skill marketplace lets you benefit from APIs discovered by the community.

← All articles

Blog

v3: The Browser Is Now the Fallback

For two years the default answer to "how does an AI agent interact with a website?" has been "spin up a headless browser." That answer is expensive, fragile, and slow. With Unbrowse v3 we are making it official: the structured API is the primary path, and the browser is what you reach for when nothing else works.

Lewis Tham

April 5, 2026

v3: The Browser Is Now the Fallback

Published 2026-04-05

What changed

v3 is a ground-up rewrite. The core thesis - that most of what agents need from the web can be resolved to clean API calls instead of pixel-scraping - has not changed, but the implementation has. The entire stack is now MCP-native, the SDK is TypeScript-first, and contributors can earn directly from the skills they publish. The formal argument is in our arXiv paper (2604.00694).

What's new

MCP native server. Unbrowse now speaks Model Context Protocol out of the box. Any MCP-compatible agent framework can connect with zero glue code.
TypeScript SDK. A typed client library for building on top of Unbrowse programmatically - resolve endpoints, execute skills, manage sessions.
Contributor dashboard. A real-time view of your published skills, earnings, and position on the leaderboard. Built to make the economics of contributing transparent.
One-command setup. git clone + ./setup and you are running. No Docker, no env juggling, no "works on my machine."
Wallet integration. Contributors connect a wallet and receive payouts proportional to skill usage. The plumbing for a contributor-owned network.
robots.txt compliance. Every resolve checks the target site's robots.txt before proceeding. Structured API access should be more respectful than browser scraping, not less.
OpenClaw plugin support. Skills can be packaged as OpenClaw plugins, opening Unbrowse's catalog to the broader agent ecosystem.

The numbers

The network has been running for months. Here is where it stands today:

162 skills in the catalog
8,760 endpoints mapped
98 domains covered
227 connected agents
1,175 total resolves
6.7x faster than headless browser
60% fewer tokens per task

The speed and token numbers come from head-to-head comparisons on the same tasks. When an agent resolves a "get the price of this product" intent through Unbrowse versus a headless browser, it completes 6.7 times faster and burns 60% fewer tokens. That gap compounds - an agent making hundreds of web calls per session sees the difference in both latency and cost.

What's next

The immediate roadmap is density: more skills per domain, more domains in the catalog, and tighter feedback loops for contributors. We are also working on multi-step skill composition - chaining resolves so an agent can complete a workflow in a single intent rather than a sequence of separate calls.

If you build agents, try it. If you want to contribute skills, the dashboard is live and the setup is one command.