The agent-to-agent communication space is crowded, but developer sam-mfb is betting that we’ve overcomplicated it. Their latest project, Kashkeshet (kkt), is a Model Context Protocol (MCP) server written in Rust that provides a bare-bones chat infrastructure for AI agents. The premise is radical in its simplicity: modern models don’t need orchestration frameworks; they just need a place to talk.

The Anti-Opinionated Approach

Unlike heavyweight orchestration tools that dictate how agents should interact, kkt offers basic primitives—channels and messages—and gets out of the way. The server listens on localhost port 7420 and requires a shared access key for security, which is intentionally minimal. It is designed for local machines or secure tailnets, relying on the agents themselves to figure out their social dynamics. As the author notes, if agents are running reasonably modern harnesses, they need surprisingly little help finding each other and coordinating tasks.

Installation and Integration

The installation process is streamlined for the hacker mindset. A single curl command installs the binary on macOS or Linux, setting up a background service that restarts on login. Connecting clients like Claude Code, Codex, or OpenCode is handled via a simple kkt connect command that writes the necessary configuration automatically. For those who prefer manual control, the MCP endpoint is exposed at http://127.0.0.1:7420/mcp. The system also includes a web interface for monitoring conversations, accessible via the generated access key stored in ~/.config/kkt/access-key.

Remote Access via Tailnet

For distributed agent swarms, kkt supports Tailscale integration. By running tailscale serve --bg 7420, users can expose the local server to other devices on their tailnet. The configuration file allows for specific host allowances, ensuring that only trusted machines can connect. However, the security model remains flat: anyone with the key and network access can see all channels, open or closed. It’s a trade-off that prioritizes ease of setup over granular permission structures.

Key Takeaways

  • Kashkeshet is a Rust-based MCP server designed for minimal agent-to-agent communication.
  • It rejects complex orchestration in favor of basic chat primitives like channels.
  • Installation is a one-liner curl command, with automatic config writing for major AI clients.
  • Security is minimal, relying on shared keys and local/tailnet network boundaries.

The Bottom Line

If you’re tired of brittle orchestration scripts, kkt’s 'let them talk' philosophy is a refreshing hack. It assumes your agents are smart enough to socialize without a chaperone, which might just be the future of multi-agent systems.