What is Unbrowse in one sentence?

Unbrowse is an open-source MCP server that lets AI agents call any website's internal APIs directly, with no per-site setup, no headless browser, and the user's real cookies for authentication.

How is Unbrowse different from Playwright or Browser Use?

Playwright and Browser Use drive a real browser for every action. Unbrowse runs a browser exactly once per site to capture the shadow APIs the page itself calls, then every later call uses those APIs directly. Across 94 live domains, Unbrowse is 3.6x faster on average and uses 40x fewer tokens than Playwright.

What is a shadow API?

A shadow API is an internal HTTP endpoint a website uses to render its own UI — for example, the JSON call that loads a product listing or a search result. Shadow APIs are not documented and not public, but they are fully functional. Unbrowse discovers them by intercepting network traffic during one normal browse session.

Do I need to write a separate MCP server for each website?

No. That is the problem Unbrowse exists to solve. One install of Unbrowse handles every site. Per-site MCP servers (a Notion MCP, a Gmail MCP, a Slack MCP, etc.) are the stack Unbrowse replaces.

Yes. Unbrowse is open source under the AGPL and runs locally. The marketplace is a shared index of captured routes; you can use it, contribute to it, and earn USDC when other agents call routes you indexed.

How do I install Unbrowse?

Run `npx unbrowse setup --mcp` once. It registers the Unbrowse MCP server with Claude Code, Claude Desktop, Cursor, Codex, Windsurf, OpenClaw, or any other MCP-compatible host. After that any agent on your machine can call any website by intent.

## Definition

What is Unbrowse?

Name: Unbrowse
Author: Unbrowse

A definition for developers, AI agents, and anyone evaluating the agentic web

## TL;DR

Unbrowse is the API layer for AI agents: one MCP server that turns any website into directly callable, cookie-authenticated API routes, so an agent never has to drive a browser through that site again.

Definition

Unbrowse is an open-source MCP (Model Context Protocol) server that lets an AI agent call any website's internal APIs directly. The first time the agent visits a site, Unbrowse runs a headless browser in the background, captures the shadow APIs the page itself uses, and publishes those routes to a shared marketplace. Every later call on the same site skips the browser entirely and returns JSON in milliseconds, signed in as the user with cookies from their real Chrome profile.

One install of Unbrowse replaces the stack of per-site MCP servers (Notion MCP, Gmail MCP, Slack MCP, Browser MCP, hand-rolled scrapers) that developers were previously wiring up one by one. The mechanism is described in the research paper Internal APIs Are All You Need; the project lives at github.com/unbrowse-ai/unbrowse.

What it is, what it is not

Unbrowse is	Unbrowse is not
An MCP server	A scraping framework you import
One install for every website	One MCP server per service you wire up by hand
Direct HTTP calls to a site's real APIs	A headless browser that clicks through the UI
Signed in with the user's real cookies	A separate auth flow per integration
Free, open source, runs locally	A hosted scraping API you pay per call

By the numbers

3.6x

mean speedup vs Playwright across 94 live domains

40x

fewer tokens consumed per task (no DOM, no screenshots)

MCP server install covers every website

Source: peer-reviewed paper, arXiv:2604.00694, n=94 domains, measured against Playwright with the same task prompts.

Frequently asked

What is Unbrowse in one sentence?: Unbrowse is an open-source MCP server that lets AI agents call any website's internal APIs directly, with no per-site setup, no headless browser, and the user's real cookies for authentication.
How is Unbrowse different from Playwright or Browser Use?: Playwright and Browser Use drive a real browser for every action. Unbrowse runs a browser exactly once per site to capture the shadow APIs the page itself calls, then every later call uses those APIs directly. Across 94 live domains, Unbrowse is 3.6x faster on average and uses 40x fewer tokens than Playwright.
What is a shadow API?: A shadow API is an internal HTTP endpoint a website uses to render its own UI — for example, the JSON call that loads a product listing or a search result. Shadow APIs are not documented and not public, but they are fully functional. Unbrowse discovers them by intercepting network traffic during one normal browse session.
Do I need to write a separate MCP server for each website?: No. That is the problem Unbrowse exists to solve. One install of Unbrowse handles every site. Per-site MCP servers (a Notion MCP, a Gmail MCP, a Slack MCP, etc.) are the stack Unbrowse replaces.
Is Unbrowse free?: Yes. Unbrowse is open source under the AGPL and runs locally. The marketplace is a shared index of captured routes; you can use it, contribute to it, and earn USDC when other agents call routes you indexed.
How do I install Unbrowse?: Run `npx unbrowse setup --mcp` once. It registers the Unbrowse MCP server with Claude Code, Claude Desktop, Cursor, Codex, Windsurf, OpenClaw, or any other MCP-compatible host. After that any agent on your machine can call any website by intent.

## Next

Read the paper: Internal APIs Are All You Need. Or install it now: npx unbrowse setup --mcp.