Auto Browser
Self-hosted browser automation for AI agents, with Docker deployment, MCP support, human takeover, and local models through Ollama, vLLM or LM Studio.
Auto Browser lets AI agents and human operators share a live Chromium browser, with a person able to take over when a login or fragile page needs attention. It's a self-hosted tool for internal dashboards, account workflows, and browser QA. The full stack runs on your own hardware through Docker Compose; Codespaces provides a hosted demo. It's open source under the MIT license.
Playwright powers browser actions, while noVNC gives operators access to the same session the agent is using. Saved auth profiles let later sessions reuse a login. Agents can inspect screenshots, page text and accessibility information, manage tabs, read downloads, and examine network activity.
MCP support connects it to Claude Code, Cursor, VS Code and Claude Desktop. A REST API is also available. For model inference, it supports OpenAI, Claude and Gemini, plus OpenAI-compatible backends such as Ollama, vLLM and LM Studio. The browser stack and stored artifacts can stay on your machine; choosing a cloud model sends inference requests to that provider.
An operator dashboard shows actions awaiting approval, including the exact arguments for governed calls. Controls include operator identity checks, site allowlists and personal-data scrubbing. Signed Witness receipts let recipients verify exported audit evidence independently of the controller.
Sessions keep separate cookies and storage, but share a browser process and takeover desktop by default. Optional per-session Docker containers separate both the browser and the human takeover surface.
