Relay for AI Agents Edge extension security report

Relay for AI Agents gives any Model Context Protocol (MCP) compatible AI agent the ability to drive a real Chromium browser. Once installed, an agent running in Claude Desktop, Cursor, Cline, Continue.dev, opencode, OpenAI Codex CLI, or any custom MCP-aware code can: • Open, navigate, and close tabs • Click, type, hover, scroll, and press keys with real input events • Capture screenshots of the viewport • Read the page's accessibility tree and find elements by CSS or text • Fill input fields, set dropdown values, upload files • Read console messages and network requests for debugging • Run arbitrary JavaScript in the page context The extension connects to a small Node.js MCP server that your agent host spawns on the same machine — or on a different machine you reach over Tailscale or another private VPN. Multiple agent sessions can share one browser at the same time. Multiple browsers can be controlled from one host by registering separate MCP entries on different ports. The popup's "Bind" feature lets you lock a browser to one specific session. Privacy: this extension transmits no data to the publisher or any third party. All browser data the agent sees stays on machines you control. No analytics, telemetry, or tracking. The source is open and auditable on GitHub. Security: for remote setups, the extension authenticates with a bearer token. We strongly recommend pairing remote use with Tailscale or another encrypted overlay network — never expose the broker port to the public internet. While the agent is controlling a tab, Chromium displays a yellow "Relay for AI Agents started debugging this browser" banner. This is shown by the browser for any extension that uses the debugger API and cannot be hidden — the trade-off is real, high-fidelity browser control instead of synthesized events that many sites detect. Consider running the extension in a dedicated browser profile. Source code, docs, and issues: https://github.com/letssetyouup/relay-for-ai-agents License: MIT