MCP and agent integrations
The MCP server uses Streamable HTTP at http://localhost:8080/mcp/, on the same port as the HTTP API. Connect a compatible client and configure Authorization: Bearer <token> when AUTH_TOKEN is set. There is no separate stdio server in the container.
Published integrations
# Claude Code
claude plugin marketplace add psyb0t/agents
claude plugin install stealthy-auto-browse@psyb0t
# Codex
codex plugin marketplace add psyb0t/agents
codex plugin add stealthy-auto-browse@psyb0t
# OpenClaw skill
openclaw skills install @psyb0t/stealthy-auto-browse
# OpenClaw MCP bridge
openclaw plugins install clawhub:@psyb0t/stealthy-auto-browse
The source skill and plugin manifests describe the integration. Set STEALTHY_AUTO_BROWSE_URL to the reachable browser URL where the integration uses it, and provide AUTH_TOKEN privately. localhost inside another container is that container, not the browser; use a shared-network service name or a reachable host address.
Dedicated tools
Single-instance mode exposes these tools from mcp_server.py:
| Purpose | Tools |
|---|---|
| Navigation and reading | goto, get_text, get_html, get_interactive_elements, screenshot, eval_js. |
| System input | system_click, system_type, send_key, mouse_move, scroll, human_pause. |
| DOM input and waits | click, fill, wait_for_element, wait_for_text. |
| Clipboard | set_clipboard for text, set_clipboard_data for rich formats, get_clipboard. |
| Events and handoff | get_events, wait_for_event, begin_handoff, handoff_status, end_handoff. |
| Challenges | detect_challenge, solve_captcha. The latter requires Anti-Captcha configuration. |
| Full action surface | browser_action, run_script. |
MCP tool names sometimes differ from HTTP action names: eval_js calls eval, and screenshot calls save_screenshot. browser_action takes an action name and a params object, so actions without dedicated tools remain available. Consult the connected tool schema for its accepted arguments and the HTTP reference for action behavior.
Multi-step work and human takeover
Use run_script to keep dependent steps in one serialized request. Atomic means no other API action interleaves with the script on that browser; it does not roll back page changes when a step fails. Script control flow supports explicit conditions and bounded loops.
For visible manual work, call begin_handoff, give the person the protected noVNC address, read handoff_status as needed, then call end_handoff to resume mutating automation. This pauses automation, not network access. It does not add viewer authentication or a temporary tunnel. Activity counts describe trusted events observed by listeners on the page where handoff began; navigation can replace those listeners.
In cluster mode, MCP exposes only run_script. HTTP still permits run_script, ping, and sleep at the top level. Put every dependent operation inside one script, or preserve the relevant session routing across requests. Use a standalone browser for interactive handoff; leaving it active across cluster scripts blocks the later script that would end it.