mozilla/firefox-devtools-mcp

What it solves

This project provides a way for AI agents (like Claude Code or Cursor) to interact with and automate the Firefox browser. It allows an LLM to perform tasks such as navigating pages, clicking buttons, filling forms, and inspecting network requests or console logs, effectively giving the AI a way to "see" and control a web browser.

How it works

It implements the Model Context Protocol (MCP) server, which acts as a bridge between an AI client and the Firefox browser. It uses WebDriver BiDi (via Selenium WebDriver) to send commands to the browser. The server can be run locally via npx or Docker, and can be configured with different "tool presets" (slim, basic, developer, mozilla) to control which browser capabilities are exposed to the AI agent.

Who it’s for

• AI Developers building agents that need to automate web browsing or perform end-to-end testing.
• Web Developers who want to use AI to debug their websites by giving the agent access to network logs, console messages, and the profiler.
• QA Engineers automating browser-based workflows using LLMs.

Highlights

• Comprehensive Browser Control: Supports page navigation, interaction (click, hover, fill, drag), and screenshots.
• Android Support: Can automate Firefox for Android via ADB.
• Flexible Tooling: Offers presets to limit or expand the agent's capabilities (e.g., removing script evaluation for security).
• Context Management: Includes a saveTo parameter for bulky output (like screenshots or large logs) to prevent overloading the AI's context window.
• Existing Session Connection: Ability to attach to a running Firefox instance to maintain cookies and logins.