mvanhorn/agent-tincan

What it solves

Personal AI agents often live in isolated environments—such as cloud VMs, sandboxes, or browser tabs—that cannot communicate with each other. This forces users to manually copy and paste information between different AI tools. Agent Tincan enables these agents to collaborate by borrowing each other's specific capabilities (e.g., one agent handles web browsing while another manages code or phone calls) without requiring complex API integrations or open ports.

How it works

The project uses a relay system built on the Tailscale network. Each agent connects to this central relay, which acts as a message broker. When one agent needs help, it sends a request to the relay, which queues the task and wakes the target agent. Because it leverages Tailscale's identity verification, agents trust each other as teammates without needing individual API keys. The system supports various platforms, including web-based agents accessed via a Chrome extension and local CLI-based agents.

Who it’s for

This is for power users and developers who manage multiple specialized AI agents across different platforms and want to create a unified, collaborative "mesh" of agents that can perform multi-step tasks autonomously.

Highlights

• Cross-Platform Collaboration: Allows agents in sandboxes, cloud VMs, and browsers to talk to one another.
• No API Keys Required: Uses Tailscale's network identity to establish trust between agents.
• Council Feature: Implements an LLM-based voting system where multiple agents answer a prompt and rank each other's responses to reach a consensus.
• Zero Inbound Ports: Agents dial out to the relay, meaning no firewall configuration or open ports are needed.
• Automated Errands: Enables agents to delegate tasks, such as having one agent book a call and another file the receipt.