Steel Browser
by steel-dev
Steel's MCP server puts a managed, self-hostable Chromium browser behind 15 MCP tools: scrape, screenshot, run multi-step sessions, and hand off to a human mid-task.
Self-host (no API key)
git clone https://github.com/steel-dev/steel-mcp-server.git && cd steel-mcp-server && npm install
About
Steel Browser is the open-source, batteries-included browser infrastructure, including session management, proxy rotation, stealth fingerprinting, and extension loading, that Steel's MCP server puts behind a standard MCP tool interface. In its default 'browse' profile the server exposes 15 tools covering everything from stateless page reads to full interactive sessions with clicking, typing, form-filling, and waiting for selectors.
A lighter 'scrape' profile strips that down to three stateless tools (steel_scrape, steel_screenshot, steel_pdf) for when an agent just needs to read pages without paying for a live session. Sessions can be handed off to a human for logins or CAPTCHAs via steel_session_handoff and resumed automatically once control returns, and the server runs either against Steel's hosted cloud with an API key or a self-hosted Steel Browser instance with STEEL_LOCAL=true.
Key features
- 15 browser-control tools in the default profile: scrape, screenshot, PDF, navigate, click/type/fill, wait-for, and more
- Stateless 'scrape' profile for simple reads with no session or billing overhead
- steel_session_handoff pauses automation for manual login/CAPTCHA handling, then resumes
- steel_batch runs multiple actions and returns a single page read to save tokens
- Works against Steel's hosted cloud (API key) or a fully self-hosted instance (STEEL_LOCAL=true)
- Session diagnostics and replay dashboards for debugging automation runs after the fact
Use cases
- Scraping pages that block plain fetch requests, returning clean markdown instead of raw HTML
- Running multi-step web workflows (login, navigate, fill a form, submit) without managing browser infrastructure
- Handing a session to a human for a login or CAPTCHA, then resuming automated steps
- Capturing screenshots or PDFs of pages as part of an automated QA or reporting pipeline
Available tools
steel_scrape
Reads a page as markdown or HTML without starting a browser session.
steel_act
Clicks, types, fills forms, selects, hovers, scrolls, or presses keys within an active session.
steel_session_create
Starts an interactive browser session for multi-step automation.
steel_session_handoff
Pauses automation so a human can take manual control, then resumes it.
steel_snapshot
Reads the page as an accessibility tree with referenceable elements instead of a screenshot.
steel_batch
Executes multiple automation steps and returns a single page read, reducing round trips.
Frequently asked questions
Do I need a Steel Cloud account to use this?
No. Set STEEL_LOCAL=true and point it at your own self-hosted Steel Browser instance instead of supplying a STEEL_API_KEY for the hosted cloud.
What's the difference between the 'scrape' and 'browse' profiles?
'scrape' exposes only three stateless read tools (steel_scrape, steel_screenshot, steel_pdf); 'browse', the default, adds the full 15-tool set for interactive sessions: clicking, typing, waiting, and handoff.