Steel MCP Server
by MCP-Mirror / steel-dev
The Steel MCP Server enables LLMs like Claude to navigate the web using Puppeteer and Steel, providing tools for standard web actions and screenshots. It's based on the Web Voyager framework and allows Claude to interact with the web.
Last updated: N/A
What is Steel MCP Server?
The Steel MCP Server is a Model Context Protocol server that allows Large Language Models (LLMs) like Claude to interact with the web. It uses Puppeteer for browser automation and integrates with Steel for browser session management. This server provides a set of tools to perform actions like navigating, searching, clicking, typing, scrolling, and taking screenshots.
How to use Steel MCP Server?
To use the Steel MCP Server, you need to configure Claude Desktop with the server details, including the command to run the server and the necessary environment variables. You can choose between running Steel in cloud mode (using a Steel API key) or in local/self-hosted mode (pointing to your local Steel instance). Follow the Quick Start guides in the README for detailed instructions.
Key features of Steel MCP Server
Browser automation with Puppeteer
Steel integration for browser session management
Visual element identification through numbered labels
Screenshot capabilities
Basic web interaction (navigation, clicking, form filling)
Lazy-loading support through scrolling
Local and remote Steel instance support
Use cases of Steel MCP Server
Searching for recipes and saving ingredients lists
Tracking package delivery status
Finding and comparing prices for specific products
Filling out online job applications
FAQ from Steel MCP Server
How do I choose between local and cloud mode?
How do I choose between local and cloud mode?
Set the STEEL_LOCAL environment variable to true for local mode and false for cloud mode. In cloud mode, you'll need a Steel API key.
What is the purpose of the numbered labels on the web page?
What is the purpose of the numbered labels on the web page?
The numbered labels identify interactive elements (buttons, links, inputs) for the click and type tools. Use these numbers to specify which element to interact with.
How do I access screenshots taken by the server?
How do I access screenshots taken by the server?
Screenshots are stored as MCP resources and can be accessed via URIs in the form screenshot://RESOURCE_NAME.
What should I do if pages are not rendering correctly?
What should I do if pages are not rendering correctly?
Try adding a delay using the GLOBAL_WAIT_SECONDS environment variable to allow pages to fully load before interacting with them.
What are the limitations of this project?
What are the limitations of this project?
This project is experimental and under active development. Session clean up isn't fully optimized, and performance may degrade after a large number of browser actions due to Claude's context window limitations.