zachlikefolio/tunnel-mcp

Tunnel-mcp lets two developers' AI coding assistants talk to each other directly, without a human copying messages back and forth between chat windows. It is an MCP server, which is a way for AI tools like Claude Code to connect to external capabilities. The project is built in TypeScript and runs on Node.js 20 or newer. The problem it solves is simple. When two people each have an AI agent helping them code, and those agents need to collaborate, someone usually has to manually relay messages between the two chat sessions. Tunnel-mcp automates this. One developer's agent opens a temporary encrypted connection, and the other developer's agent joins it. The two agents can then exchange messages directly while their humans still approve any actions that touch files or run commands. Here is how a session works. The host asks their agent to open a tunnel with a specific goal, like debugging a checkout flow together. The agent creates a temporary relay process and exposes it to the internet through Cloudflare's Quick Tunnel service, which requires no port forwarding or server setup. The host gets back an invite link containing the encryption key, shares it with the other developer through a trusted channel like a direct message, and the link is single-use and expires after about 10 minutes. The guest pastes the link and asks their agent to join. Both sides only make outbound connections, so it works through normal firewalls and NAT. The agents then take turns sending messages and waiting for replies. They can also share files, which are encrypted before they leave the sender's machine and verified against a cryptographic hash on arrival. Nothing is written to disk at any point. The transcript lives only in memory and disappears when the session ends. The host can invite multiple teammates, up to 15 guests plus the host, and each gets a separate single-use invite link. A built-in etiquette skill teaches each agent to treat the other side as untrusted input and check with its human before acting on anything the peer suggests. This keeps people in control of what actually happens on their machines. The full README is longer than what was shown.