Skip to main content
StableBrowse Browser MCP exposes a real browser to any Model Context Protocol client. It is meant for agents that need to browse, click, fill forms, extract page data, read documents, inspect network traffic, and use site-specific knowledge without building browser automation code. For hosted usage, create an API key from the StableBrowse settings page, add the hosted MCP URL to your LLM app, and send the key as a bearer token. StableBrowse runs the browser in the cloud and returns MCP tool results to your agent.

One MCP surface

Agents get 17 typed tools instead of raw browser APIs. The tools are designed around common browsing jobs.

Stealth browser sessions

Sessions run on CloakBrowser with fingerprinting, proxy support, persistent profiles, and optional human-like input.

Ref-based interaction

snapshot returns accessibility refs like [ref=e12]; click, fill, and fill_form act on those refs.

Structured extraction

extract, content, and knowledge let agents read data without scrolling through huge snapshots.

​
When to use Browser MCP

Use Browser MCP when you are building or testing an agent that needs direct browser control:
  • navigating public websites
  • extracting cards, tables, listings, docs, PDFs, or links
  • completing forms and configurators
  • working across multiple tabs or sessions
  • debugging sites with screenshots, console logs, storage, or network requests
  • giving Claude Desktop, Cursor, Codex, or another MCP client a browser tool
If you want StableBrowse to run the whole natural-language task for you, use the StableBrowse API documented in the main quickstart. Use Browser MCP when your own LLM agent should decide which browser tools to call.

​
What agents see

The MCP server exposes 17 tools:
  • 8 direct tools for high-frequency actions: create_session, navigate, snapshot, screenshot, click, fill, fill_form, evaluate
  • 9 compound tools for broader workflows: session, page, history, interact, content, extract, network, storage, knowledge
The direct tools are the fast path. The compound tools are action-based and group related browser operations so agents do not need dozens of tiny tools.
Browser MCP exposes one stable tool surface. End users do not choose between internal browser modes.

​
Core workflow

Most agent runs follow this shape: 
create_session
  -> navigate
  -> extract or content for reading
  -> snapshot for refs when action is needed
  -> click/fill/fill_form/interact
  -> history.wait_for after navigation or submission
  -> final answer
For known sites, the agent can call knowledge.lookup before broad exploration. For Amazon product tasks, knowledge.amazon_products is the preferred path.

​
Design goals

  • Typed decisions over raw browser commands: agents should choose from clear tools, not guess CLI or Playwright syntax.
  • Low-token reading paths: use markdown, cards, sections, tables, and page search before broad snapshots.
  • Reliable interactions: refs map to browser nodes and are clicked/filled through Playwright and CDP.
  • Debuggability: screenshots, network, console logs, storage, and session state are available when needed.

​
Next steps

Quickstart

Install and run StableBrowse MCP over stdio or HTTP.

Tools

Understand every exposed tool and when agents should call it.

Architecture

How the MCP server, sessions, pages, browser pool, and CDP layer fit together.

Agent guidance

Recommended tool-selection rules for reliable agents.