Install
https://mcp.browserbase.com/mcpWho it is for
If you want an LLM to open pages in a cloud browser, observe actionable elements, or extract information, this service fits. Officially it uses Browserbase and Stagehand to provide cloud browser automation, and it can interact with pages, extract information, and run automatic actions.
Do not treat this current GitHub repository as a maintained production implementation: the excerpt says the repository is archived and no longer maintained, historical reference only, and it should not stand for Browserbase's current production service. The main path is hosted MCP, not cloning this archived code on the machine.
The current excerpt names six tools: start, end, navigate, act, observe, extract. The first time, create a session then observe; do not act first. The cloud browser will visit the URL you give and may incur session charges; do not treat it as having no network or no privacy boundary.
First try in 3 minutes
Officially, connect the hosted service with SHTTP to use the full hosted capacity; do not also run STDIO self-hosting in parallel:
https://mcp.browserbase.com/mcp
Point an SHTTP-capable MCP client at https://mcp.browserbase.com/mcp . The excerpt says this is the easiest start, and they cover Gemini's LLM cost. The excerpt does not lock a client config filename.
After connecting, do one low-risk try in this order: first call start (no input) to create or reuse a Browserbase session. Then call navigate, filling url with an address you already meant to open; do not invent a sample URL that is not in the excerpt. After the page opens, call observe, writing in instruction the actionable elements you want to observe. Do not call act on the first try.
Side effects: start creates or reuses a cloud session; navigate makes that session open the given URL. Confirm the URL can be opened before submitting. The success signal is observe returning an observation of actionable elements on the page, not a failure to reach https://mcp.browserbase.com/mcp , or a session that never started.
Use cases
Case 1: observe what can be clicked on the page. The excerpt says observe is for Observe actionable elements on the page, input { instruction: string }. Call it after start and navigate. On success you see actionable elements given by instruction. Exact return fields are not written into the excerpt.
Case 2: extract data from the page. The excerpt says extract is for Extract data from the page, input { instruction?: string }. On success you should get an extraction result, not an observation list. instruction may be omitted, but the excerpt does not explain the extract range when omitted; do not treat the default behavior as proven.
act will Perform an action on the page, input { action: string }. This is a write-oriented page action; first confirm action is the step you want. end closes the current session; you can close it when done; the excerpt writes no undo or replay.
Boundaries
The main path is hosted SHTTP: https://mcp.browserbase.com/mcp . Supported transports are STDIO and SHTTP; SHTTP is recommended. This repository is archived. The excerpt also gives self-hosted npx @browserbasehq/mcp, and shows environment variables BROWSERBASE_API_KEY, BROWSERBASE_PROJECT_ID, GEMINI_API_KEY; that is a competing path, so do not take it on the first try, and do not fill any example keys.
The product page is https://www.browserbase.com/ . The docs entry is https://docs.browserbase.com/integrations/mcp/introduction . The official repository is https://github.com/browserbase/mcp-server-browserbase . Do not invent selector parameters, session billing numbers, or undo steps that never appear.