onbridge

Browser control for AI agents via MCP

OnBridge connects AI agents running on your own computer to the browser you already use, with your sessions, your logins and your extensions, over the Model Context Protocol (MCP). It differs from a headless automation library in three ways that matter. REAL INPUT. Clicks and keystrokes are dispatched through the Chrome DevTools Protocol, so they arrive as genuine trusted events. Synthetic DOM events are rejected by native form submission, drag-and-drop, canvas apps and most modern web applications. A GOVERNANCE LAYER. Actions that spend money, delete things or read credentials are held for your approval, and approvals fail closed. Page content is fenced as untrusted so a hostile page cannot instruct the agent. You choose the tab scope, restrict the agent to named domains, and can pause it at any moment. Control Mode revokes itself after 30 idle minutes. AN AUTHENTICATED, ENCRYPTED CHANNEL. The extension talks only to a process on your own machine over loopback. Pairing uses an ECDH key exchange, every frame is AES-256-GCM encrypted, and web pages are refused by origin. Nothing is sent to any server we operate, because there is none. Works with any MCP-capable agent: Claude Code, Codex, Cursor and others. SETUP. This extension needs the OnBridge MCP server running on your machine; install it from the repository linked below. Then open the side panel, turn on Control Mode and approve the agent once. Every later session connects silently. Free and open source under GPL-3.0. Source, issues and documentation: https://github.com/prakersh/onbridge