Browserbase MCP Server with Stagehand for LLM Automation

Overview

Browserbase MCP Server is a self-hostable implementation of a Model Context Protocol (MCP) server that connects large language models to a cloud browser automation stack. It combines Browserbase (cloud browser sessions) with Stagehand (LLM-driven browser automation) so LLMs can open pages, click and type, observe actionable elements, and extract structured data from websites. The server follows the MCP spec, exposing a small set of tools LLMs can call as part of a broader AI application.

There are two deployment modes: use Browserbase’s hosted MCP endpoint for the fastest setup, or run the server locally (STDIO/SHTTP) to keep everything on-premises. The hosted SHTTP option includes managed LLM access (Stagehand models) and is recommended for rapid adoption; the local server is useful for integration testing, custom models, or air-gapped environments.

Features

• Exposes MCP-compatible tools to control cloud browser sessions from LLMs
• Integrates Browserbase cloud browsers with Stagehand LLM automation
• Supports two transport modes: SHTTP (remote HTTP MCP) and STDIO (local MCP)
• Start/end browser sessions, navigate, perform actions, observe elements, and extract data
• Configurable viewport, persistence, proxy/stealth options, and Stagehand model selection
• Works with Puppeteer/Playwright-backed browser instances via Browserbase

Installation / Configuration

Two recommended approaches: use the hosted MCP server (SHTTP) or run the server locally (STDIO). Set the following environment variables when running the server (or provide via your process manager):

• BROWSERBASE_API_KEY — your Browserbase API key
• BROWSERBASE_PROJECT_ID — Browserbase project id
• GEMINI_API_KEY (optional) — API key for Google Gemini if using that model

SHTTP (Hosted, easiest)

• Add the hosted MCP endpoint to your MCP client config to point at Browserbase’s managed server.

Example client config (SHTTP):

{
  "mcpServers": {
    "browserbase": {
      "type": "http",
      "url": "https://mcp.browserbase.com/mcp"
    }
  }
}

If your client cannot use SHTTP directly, a bridge invocation via npx is possible:

{
  "mcpServers": {
    "browserbase": {
      "command": "npx",
      "args": ["mcp-remote", "https://mcp.browserbase.com/mcp"]
    }
  }
}

STDIO (Self-hosted)

• Install from npm or clone this repository and run locally or in Docker.

Run via npx (recommended for local use):

{
  "mcpServers": {
    "browserbase": {
      "command": "npx",
      "args": ["@browserbasehq/mcp"],
      "env": {
        "BROWSERBASE_API_KEY": "",
        "BROWSERBASE_PROJECT_ID": "",
        "GEMINI_API_KEY": ""
      }
    }
  }
}

Run from source:

git clone https://github.com/browserbase/mcp-server-browserbase.git
cd mcp-server-browserbase
npm install
npm run build
node cli.js

Run in Docker:

# build image
docker build -t mcp-browserbase .
# example docker run in MCP client config
{
  "mcpServers": {
    "browserbase": {
      "command": "docker",
      "args": ["run","--rm","-i","-e","BROWSERBASE_API_KEY","-e","BROWSERBASE_PROJECT_ID","-e","GEMINI_API_KEY","mcp-browserbase"]
    }
  }
}

Available Tools

This MCP server exposes six primary tools LLMs can call. Each tool expects JSON input as shown.

Stagehand is used under the hood to interpret freeform instructions (act/observe/extract) into concrete browser operations.

Configuration Flags

Common runtime flags accepted by the server:

Pass flags via the args array in your MCP client config or as CLI parameters when launching the server.

Use Cases

• Data extraction / web scraping: Ask an LLM to open a product page, extract price, name, and availability into JSON using extract.
• Form automation: Instruct the LLM to fill and submit multi-step forms (login, checkout) with act and navigate calls.
• Research assistants: Have an LLM browse a sequence of pages, summarize findings with observe, and return citations.
• End-to-end testing: Automate UI flows via Stagehand scripts driven by LLM-generated instructions and validate visible outcomes.
• Augment chatbots: Integrate live web data into a chat agent by letting the model fetch and parse pages on demand.

Links & Resources

• Repository: https://github.com/browserbase/mcp-server-browserbase
• Model Context Protocol: https://modelcontextprotocol.io/introduction
• Browserbase docs and Stagehand examples are helpful when authoring extraction/observation instructions.

This server is geared toward developers building LLM integrations that need reliable, programmable browser context and a compact set of MCP tools to orchestrate browsing tasks.