Invisible Playwright MCP lets coding agents browse sites that block bots

A coding agent sent to book a flight, check a dashboard, or copy data from a web page is stopped by the site's captcha before it can finish.

Invisible Playwright MCP is a Model Context Protocol server that drives a stealth Firefox instead of the standard Playwright browser, so an AI agent can finish the job. The repository feder-cr/invisible_playwright_mcp is MIT-licensed, written in Python, and carries 2,645 stars in the nine days since it was created on 29 September 2026. The README's own one-line pitch: "Other AI browser agents get captchas. This one is invisible to anti-bots."

How it is different from the stock Playwright MCP

Microsoft's official Playwright MCP ships a Chromium browser with no stealth, so sites that run bot checks see it and block it. The invisible variant wraps two libraries from the same author, invisible_playwright and invisible_core, which handle fingerprint management, mouse and keyboard timing, and proxy support. The README describes the behaviour as driving a page "the way a person would: the pointer moves, keys are pressed."

The project exposes the usual Playwright-style tools (navigate, click, type, screenshot, extract) to an agent over MCP, so Claude Code, Codex and Gemini CLI can call them with no new driver code. A standalone web UI is also included for use with an OpenRouter key, which is useful for anyone who wants the browser without connecting it to an IDE.

What the server adds to an existing agent setup

The server installs as a plugin in Claude Code, Codex and Gemini CLI, so an existing agent setup gets browser tools that work on captcha-protected sites with no new driver code to write. HTTP and SOCKS5 proxy support, a seed for a reproducible browser identity, and persistent browser profiles are configuration options.

The example in the README is a flight search across many dates on an airline site that uses calendar widgets rather than an API: the kind of task a person would use a browser for, where "a browser rather than an API, and a person's judgement about what is on the page" is the point.

What the author asks of users

The README includes a short responsibility paragraph, worth quoting because the tool is dual-use: "Read the terms of the sites you point it at, respect their rate limits, and do not submit anything a human has not read." The project is a browser the user already controls, so the author treats consequences of its use as the user's problem.

Data flow is local except for three expected destinations: the sites visited, the model provider (OpenRouter if the standalone UI is used), and GitHub for engine downloads. There is no proprietary server between the agent and the sites it opens.

Source

• feder-cr/invisible_playwright_mcp, the MCP server on GitHub.