Set up Steel with an AI agent

Full docs index: https://docs.steel.dev/llms.txt

Steel Documentation

Steel is the open-source browser API for AI agents — managed cloud browsers with stealth, residential proxies, CAPTCHA solving, persistent profiles, session replays, and agent observability. Use Steel to launch cloud browsers, scrape content, and automate web tasks.

Quick Reference

• Install: npm install steel-sdk (Node.js) or pip install steel-sdk (Python)
• CLI: curl -sSf https://setup.steel.dev | sh
• Auth header: steel-api-key: <your-key>
• Auth env var: STEEL_API_KEY
• API base URL: https://api.steel.dev
• WebSocket: wss://connect.steel.dev?apiKey=<key>&sessionId=<id>
• MCP server: https://mcp.steel.dev/mcp (OAuth sign-in, or Authorization: Bearer <your-key>)
• API reference: https://steel.apidocumentation.com/api-reference
• Brand design language (DESIGN.md): https://docs.steel.dev/DESIGN.md

Agent Instructions

• For the simplest one-liner scrape, use the CLI:

  steel scrape https://example.com
  

• For simple scraping without a browser session, use the REST scrape endpoint:

  curl -X POST https://api.steel.dev/v1/scrape \
    -H "steel-api-key: YOUR_KEY" \
    -H "Content-Type: application/json" \
    -d '{"url": "https://example.com"}'
  

• For browser automation, connect Puppeteer or Playwright via WebSocket:

  import Steel from 'steel-sdk';
  import puppeteer from 'puppeteer-core';
  
  const client = new Steel({ steelAPIKey: process.env.STEEL_API_KEY });
  const session = await client.sessions.create();
  const browser = await puppeteer.connect({
    browserWSEndpoint: `wss://connect.steel.dev?apiKey=${process.env.STEEL_API_KEY}&sessionId=${session.id}`,
  });
  // ... use browser ...
  await browser.close();
  await client.sessions.release(session.id);
  

• For agent-style browser control over plain HTTP, use the Commands API. Take a snapshot, then act on the refs in it:

  curl -X POST https://api.steel.dev/v1/sessions/SESSION_ID/commands \
    -H "steel-api-key: YOUR_KEY" \
    -H "Content-Type: application/json" \
    -d '{"commands": [{"command": "navigate", "url": "https://example.com"}, {"command": "snapshot", "interactive": true}]}'
  

  Then send {"after": "<next from the response>", "commands": [{"command": "click", "target": "e1"}]}. Docs: https://docs.steel.dev/overview/commands-api/overview
• To give an MCP client Steel browsers, add the hosted MCP server. In Claude Code:

  claude mcp add --transport http steel https://mcp.steel.dev/mcp
  

  The client signs in with OAuth on first use. Docs: https://docs.steel.dev/overview/mcp
• Python SDK:

  from steel import Steel
  client = Steel(steel_api_key="YOUR_KEY")  # or set STEEL_API_KEY env var
  result = client.scrape(url="https://example.com")
  print(result.content.html)
  

• client.scrape() returns a ScrapeResponse with:
  • result.content.html — full HTML string
  • result.content.markdown — markdown version
  • result.content.cleaned_html — cleaned HTML
  • result.content.readability — readability text
  • result.metadata.status_code — HTTP status (int)
  • result.metadata.title — page title
  • result.links — list of extracted links
• Always release sessions when done: client.sessions.release(sessionId)
• Do NOT use session.websocketUrl directly — construct the WSS URL as shown above
• The Node SDK constructor param is steelAPIKey (not apiKey)
• The Python SDK constructor param is steel_api_key (not api_key)
• The Python package installs as pip install steel-sdk but imports as from steel import Steel
• sessions.create() accepts an optional sessionId (Node) / session_id (Python) UUID when you need the ID before the session exists; omit it and Steel generates one
• The Python sessions.create() session timeout param is api_timeout (not timeout, which is the HTTP request timeout)
• Any docs page is available as markdown by appending .md to its URL, for example https://docs.steel.dev/overview/steel-cli.md; AI user agents (Claude, Cursor, GPT) receive markdown automatically at the canonical URL
• /llms-full.txt concatenates every page into one file (large, roughly 230k tokens); prefer fetching the individual .md pages you need over reading the whole bundle