SuperAPI AI Directory
Partner with us中文AI newsFavorites
All toolsAI agentsinvisible_playwright_mcp

invisible_playwright_mcp

invisible_playwright_mcp is a Playwright-based MCP server that lets AI agents browse the web on an anti-detection stealth Firefox browser, reportedly evading anti-bot mechanisms and CAPTCHAs. It offers two ways to use it: connect to an assistant via MCP, or use the standalone Web UI (bring your own OpenRouter key). It supports options such as proxies, profiles, and seeds, and can be used for web automation, scraping, and computer-use tasks.

Open website

About this tool

Purpose

invisible_playwright_mcp is a Playwright MCP server whose goal is to let AI agents browse the web on an anti-detection stealth Firefox browser, reportedly invisible to anti-bot mechanisms and CAPTCHAs. It is aimed at tasks that require real web automation (rather than APIs) and that require a human to judge page content, such as web scraping, browser automation, and computer use.

Main capabilities

  • Drives the web with a stealth Firefox engine, operating in a human-like way: pointer movement, key input.
  • Offers two ways to use it:
    1. Connect from an assistant via MCP (after installing with PowerShell on Windows, install it in the assistant).
    2. Standalone use: comes with a Web UI, the user provides an OpenRouter key, chat on the left, live browser on the right, visit http://127.0.0.1:8765。
  • Supports options such as proxies, profiles, and seeds, used to control browser identity and network egress.
  • Comes with a wiki, including an AI browser agent guide, checklists for when blocked, articles about OpenAI Operator, etc.; articles/ contains examples, records, and their outputs.
  • The same series also includes: invisible_playwright (the engine in the form of a Python library, with Playwright's API), invisible_core (from seed to fingerprint, preferences, proxy, and geolocation).

Usage

  • MCP server: invisible_playwright_mcp with no subcommand is the server; invisible-playwright-mcp ui is the interface.
  • Command-line options:
    • --openrouter-key: your key, or use the OPENROUTER_API_KEY variable.
    • --model: OpenRouter model id, or INVISIBLE_MCP_MODEL, default z-ai/glm-5.3-flash.
    • --proxy: optional, e.g. http://user:pass@proxy.example.com:8080 Or socks5://proxy.example.com:1080, both host and port are required, and the time zone, region, and exit change accordingly.
    • --binary: Use an existing engine binary; it must be a build pinned by seal, otherwise startup is refused; this skips the download but does not skip the version check.
    • --seed: An integer; the same seed produces the same browser identity on every run.
    • --profile-dir: The directory where the profile is saved, so that logins and cookies persist after a restart.
    • --headed: Show the browser window.
    • --host, --port: Defaults to 127.0.0.1 and 8765; changing the host exposes the interface without authentication.
  • .env: You can place a .env file in the run directory to write secrets and the binary path; it is read at startup, and the precedence is --flag > environment variables > .env > defaults; only the current directory is read, with no upward search; the startup line lists the variable names that are applied but does not print the values.
  • Passing the key via --openrouter-key leaves the key in shell history and, on Linux, it also appears in the process list; using environment variables or .env avoids this.

Use Cases

  • Web automation tasks that require a real browser rather than an API.
  • Scenarios that require checking page content item by item and making judgments, such as checking the lowest ticket price day by day (example given in the material: Milan to Lisbon, economy class, one checked bag, one adult, checking day by day from the 12th to the 16th of next month; the calendar control requires clicking dates rather than typing, unavailable dates must be stated, and numbers must not be guessed).
  • Web scraping and computer-operation tasks.

Limitations and Notes

  • The tool automates a browser controlled by the user; you should read the target website's terms, comply with rate limits, and not submit content that has not been read by a human.
  • The browser agent runs on the user's machine and has no server of its own. Data flows: the websites visited see the browser (just like any Firefox); the model provider (the Web UI sends the conversation and the page content read by the agent to OpenRouter, using your key; with MCP, the connected client uses its model to do the same); GitHub (on first startup of the server or interface, the engine is downloaded from a GitHub release, and when a proxy is set, the GeoIP database is downloaded; on every browser startup it also downloads from GitHub release a one-line counter file used to count startups; this request carries no identifier or your information, and GitHub will see the IP address that is shown for any HTTPS request).
  • Apart from this, no other data is collected, and nothing is sent to the author. Sessions, profiles, and screenshots are stored locally (under INVISIBLE_MCP_HOME if it is set, otherwise in the system application data directory), can be deleted by you, and are not retained anywhere else.
  • License: MIT. All content distributed before September 2, 2026 is released under AGPL-3.0 and remains under that license.
  • More details about charges, pricing, platform support scope, and similar matters are not yet provided in the material.

Related tools

Browse category