steel-dev/steel-mcp-server

Give Claude, Cursor, or any MCP client a real Chrome in the cloud. Read pages that block a plain
fetch, fill forms, take screenshots, and hand the browser to a person when a login or a CAPTCHA
needs one. The browser is run by Steel.
Status: 3.0.0. Download it from Releases.
It is not published to npm. mcp.steel.dev is not live yet.
## Install
You need a Steel API key. The free tier is enough to try it.
### Claude for macOS or Windows
Download steel-mcp-2.0.1.mcpb
and open it. Claude installs it and asks for your Steel API key. No Node install is needed.
When you replace an installed bundle, quit and reopen Claude after the installer finishes, then start
a new conversation so Claude reads the new tool catalog.
### Claude Code
bash
claude mcp add steel -e STEEL_API_KEY=your-steel-api-key -- npx -y github:steel-dev/steel-mcp-server
The first run clones and builds the server, which takes about a minute. Later runs start at once.
### Cursor, VS Code, and other MCP clients
Add this to the client's MCP configuration:
json
"mcpServers": {
"steel": {
"command": "npx",
"args": ["-y", "github:steel-dev/steel-mcp-server"],
"env": { "STEEL_API_KEY": " " }
To run from a checkout instead, npm install builds the server; point command at node and
args at /absolute/path/to/steel-mcp-server/dist/stdio.js.
### Self-hosted steel-browser
Run the steel-browser image and set STEEL_LOCAL=true
instead of an API key. No key is needed or sent. Self-hosted Steel runs one browser session at a
time and has no managed proxies, browser profiles, managed credentials, or CAPTCHA solving; a tool
that asks for one of those gets a named explanation instead.
### A shared endpoint for a team
dist/hosted.js serves the same tools over Streamable HTTP, with each caller bringing their own
Steel key. Claude Code connects to it directly; Claude Desktop needs the mcp-remote bridge.
Deployment, Redis, tracing, and the client snippets are in docs/HOSTING.md.
## What you can ask
| Ask | What happens |
| "Read this page and summarize the pricing table." | One steel_scrape. No browser session, nothing to release |
| "Compare the price of this product across these three shops." | Three stateless reads, or a session when a shop needs JavaScript to render |
| "Sign in to my account and check last month's invoice." | A session, a snapshot, and the browser handed to you at the login wall. The server never guesses a password |
| "Fill out this application with the details from my CV." | The agent fills the ordinary fields, then steel_session_handoff lets you choose the CV locally in the same browser |
| "Screenshot the top of this article for a slide." | One steel_screenshot, shown inline when small enough and always linked for download |
| "What happened in my last browser session?" | steel_session_diagnostics reads the latest released session. No new browser is started |
## How it works
Read first, browse when you must. steel_scrape answers most questions about a page and starts
no billed session. A session is created only to interact, and released when the task is done.
Elements, not pixels. steel_snapshot and steel_find read the page as an accessibility tree
and give targetable elements @eN references. steel_act clicks, types, and fills by reference.