yoda.digital / open source

Cluster · Agent infrastructure

Coordination and control for AI agents

Two projects that make AI agents usable at scale: mycelium for end-to-end encrypted messaging between Claude Code instances, and mcp-gitlab-server for full GitLab control from an agent. Together they form the base layer of an agent fleet that collaborates.

Related projects

Why this matters

A single agent is a toy. A fleet of agents that collaborate with each other and operate real systems is infrastructure. The base layer: how agents authenticate each other, how instructions move without plaintext, how they hook into existing code systems.

mycelium ships E2E messaging between Claude Code instances: ephemeral Curve25519 keys for forward secrecy, Ed25519 identity pinning (TOFU + STS), replay protection with a 30-minute expiry, and a stateless relay that can multi-location failover. Works over SSH, tunnels, or the public internet.

mcp-gitlab-server is the expansion path: 86 tools covering projects, MRs, CI/CD, releases, wikis, members, and activity. Read-only mode keeps AI context safe. Combined, they let an agent fleet plan, write code, open PRs, and stay in sync without human supervision.

How to start

  1. 01
    For GitLab control: npx -y @yoda.digital/gitlab-mcp-server (configuration on the install page).
  2. 02
    For agent-to-agent messaging: clone mycelium and start the relay with bun install && bun run relay.ts (npm publish coming soon).
  3. 03
    Start with a simple case: two Claude Code instances on the same machine exchange status updates over mycelium.
  4. 04
    Extend to cross-host only once you understand the identity model — TOFU only works if the first connection is trusted.

Both projects are written for real use, not demos. Mycelium is moving to npm under @yoda.digital/mycelium; the portal shows the “Coming soon to npm” badge until publish is official.

See also