Browser MCP Server

tl;dr: there is no general “Anthropic browser MCP” to install. The old @anthropic/mcp-server-browser package never existed, and Anthropic’s modelcontextprotocol/servers-archived/puppeteer is read-only since 2025-05-29. The replacement most teams pick is Microsoft’s Playwright MCP (@playwright/mcp@latest, 36.9k stars). For driving an already-logged-in Chrome, BrowserMCP/mcp (formerly browsermcp/mcp, 7.1k stars) is the lighter-weight option.

Pick one

Install — Playwright MCP (default)

@playwright/mcp@latest is what the Playwright MCP page goes deep on. The shortest install:

{
  "mcpServers": {
    "playwright": {
      "command": "npx",
      "args": ["@playwright/mcp@latest"]
    }
  }
}

Microsoft’s README: “drives pages from the accessibility tree, so the model does not need vision.” It also says coding agents may prefer Playwright CLI + skills because MCP tool schemas are heavy on tokens.

Install — Browser MCP (BrowserMCP/mcp)

A simpler alternative that just automates your existing Chrome via the DevTools Protocol:

{
  "mcpServers": {
    "browser": {
      "command": "npx",
      "args": ["-y", "browsermcp"]
    }
  }
}

Useful when the model needs access to a browser session that’s already logged in to whatever you’re testing.

Anthropic’s archived Puppeteer server

The Anthropic reference server is at modelcontextprotocol/servers-archived/src/puppeteer. It’s read-only and unmaintained. The npm package still installs but you’ll get stale Puppeteer. Don’t use it for new projects — pick @playwright/mcp@latest instead.

Gotchas

• The Microsoft MCP has a browser_run_code_unsafe tool in its list. Read their Security section before enabling page-JS evaluation from the model.
• Page snapshots via the accessibility tree are not screenshots — if you need visual review, take a screenshot explicitly.
• Browser automation over MCP stacks a lot of token overhead per call. If you’re driving a deterministic script, prefer the Playwright CLI or a one-off Node script.

Alternatives in the registry

The awesome-mcp-servers Browser Automation section has ~30 entries; notable forks of Playwright:

• browserless/browserless-mcp — headless scraping infra
• firecrawl/firecrawl-mcp-server — firecrawl_interact lets the model click before extracting (JS-heavy / auth-gated sites)
• automatalabs/mcp-server-playwright — community Playwright wrapper
• executeautomation/mcp-playwright — same idea, more tools

Related

• Playwright MCP — full deep dive
• MCP Servers Explained