LvcidPsyche/auto-browser

Give your AI agent a real browser — with a human in the loop. Open-source MCP-native browser agent.

What it solves

Auto Browser provides a way for AI agents to interact with a real web browser while allowing humans to remain in the loop. It solves the problem of brittle web automation by enabling human takeover of live sessions, managing reusable authentication profiles so agents don't have to log in repeatedly, and providing safety rails like approval gates and PII scrubbing for authorized workflows.

How it works

The system consists of a controller API and a browser node running Chromium via Playwright. It is built as an MCP (Model Context Protocol) server, allowing it to integrate directly with LLM clients like Claude Desktop or Cursor. It supports a wide range of model providers (OpenAI, Claude, Gemini, and OpenAI-compatible endpoints like Ollama or Groq) to drive the browser. A noVNC interface allows human operators to visually monitor or take over the session in real-time.

Who it’s for

  • Developers building MCP-powered agent workflows that require a full browser rather than simple HTML fetching.
  • Teams managing internal dashboards or admin tools where human recovery is needed for brittle sites.
  • QA engineers performing operator-assisted browser debugging.
  • Users who need to automate account workflows with persistent, reusable login sessions.

Highlights

  • MCP-native: Integrated as an MCP server from the start for seamless LLM agent connectivity.
  • Human-in-the-loop: Live session takeover via noVNC when automation fails or requires human intervention.
  • Auth Profile Management: Ability to save and reuse named authentication states to bypass repeated logins.
  • Safety and Compliance: Built-in approval gates, operator identity tracking, PII scrubbing, and policy presets (e.g., strict or balanced).
  • Broad Model Support: Compatible with almost any frontier model via OpenAI-compatible adapters.
  • Convergence Harness: A tool for "Agent Skill Induction" that records and verifies browser traces to create repeatable agent skills.