Puppeteer MCP Server

Reference MCP server that lets LLMs drive a real Chromium browser via Puppeteer.

Free· Free and open source (MIT license). No hosted service; you run it locally under Node/Docker.MCP Servers

Best for

Developers prototyping MCP-based agents on their own machine who need an LLM to see and interact with JavaScript-rendered pages without building a bespoke browser harness.

Skip if

Production scraping of anti-bot-protected sites, multi-tenant SaaS agents, or teams that need an actively maintained, sandboxed, session-aware browser tool.

Puppeteer MCP is one of the original reference Model Context Protocol servers published by Anthropic in the modelcontextprotocol/servers repository. It exposes a headful or headless Chromium browser to any MCP-aware client (Claude Desktop, Claude Code, Cursor, VS Code Copilot Chat, Zed, and other MCP hosts) through a small, focused set of Puppeteer-backed tools: puppeteer_navigate, puppeteer_screenshot (full-page or element-scoped, returned as base64 PNGs and stored as named resources), puppeteer_click, puppeteer_hover, puppeteer_fill, puppeteer_select, and puppeteer_evaluate for arbitrary in-page JavaScript. Console logs from the running page are also exposed as an MCP resource so the model can read what the browser is emitting. It is aimed at developers and prompt engineers who want to give an LLM eyes and hands on the live web without wiring up a bespoke browser-automation harness. Typical workflows are ad-hoc research on JavaScript-rendered pages, filling out and submitting forms, capturing visual evidence for agents that write bug reports, scraping a specific element the model can identify on the fly, and letting an agent smoke-test a local staging site. Installation is a single npx or Docker command wired into the host's mcp config; the NPX flavor pops a real Chromium window (useful when you need to sign in interactively), while the Docker flavor runs headless Chromium in a container. Note that as of 29 May 2025 the repo was moved to modelcontextprotocol/servers-archived and marked read-only, so it is now a frozen reference rather than a living project.

Editor's take

The obvious reference implementation for anyone learning MCP: read the source, run it once, understand the pattern, then move on. It's genuinely useful for local experiments, but the May 2025 archival is a real signal - for anything you plan to ship, look at the Playwright MCP server or one of the hosted browser-MCP providers that picked up where this left off.

— The AI Tool Bible editorial team

Pros

✅ Zero-config bridge between any MCP client and a real Chromium instance, including JS-heavy SPAs that curl or fetch-based tools cannot see.
✅ Complete open-source TypeScript reference implementation, easy to fork as a starting point for a hardened production browser-MCP.
✅ Ships both NPX (headful, interactive login friendly) and Docker (headless, sandboxed) install paths.
✅ Screenshots are exposed as MCP resources with stable names, so a model can refer back to earlier captures across a conversation.
✅ puppeteer_evaluate gives the model an escape hatch to run arbitrary in-page JS when the fixed click/fill/select tool set is not enough.
✅ Console log resource lets the agent read runtime errors without a separate devtools bridge.

Cons

⚠️ Archived on 29 May 2025 (read-only) and no longer maintained by Anthropic; expect Puppeteer/Chromium drift over time.
⚠️ Runs a browser on your machine with access to local files and localhost/private IPs, which is a real prompt-injection and SSRF risk if the model visits hostile pages.
⚠️ No built-in proxy rotation, stealth patches, CAPTCHA handling, or Cloudflare-turnstile bypass, so it is not suitable for anti-bot-hardened targets.
⚠️ Selector-based tools rely on the model producing correct CSS selectors; brittle on frequently-changing DOMs.
⚠️ No session persistence, cookie import, or multi-tab management primitives out of the box.
⚠️ Community forks (Playwright MCP, browserless MCP, hyperbrowser, etc.) have moved past this reference in features and active support.

Use cases

Give Claude a browser to research JS-rendered pagesHave an agent fill and submit a web formTake element-scoped screenshots for bug reportsSmoke-test a local staging site from an LLMScrape specific fields the model identifies visuallyAutomate a signed-in web workflow via headful loginExtract data with in-page JS via puppeteer_evaluatePrototype a custom browser-MCP by forking the reference