DEV Community

Rumblingb
Rumblingb

Posted on • Originally published at smithery.ai

Spin Up a Multi‑Machine MCP Server Mesh with Cord in 10 Minutes

#ai #mcp #devtools #agentpay

Hook:

Building a distributed AI agent stack feels like juggling flaming chainsaws: you need fast discovery, secure auth, and zero-copy data sharing. If you’ve only ever spun up a single MCP server locally, you’ll thank me for showing you how to wire three of them into a seamless mesh in under ten minutes—all without writing a single line of glue code.

  1. Pick a catalog entry

    Grab a ready-made implementation from Cypress Creek’s catalog:

    https://smithery.ai/servers/vishar-rumbling. It ships with a minimal MCP server, Cord agents, and a lightweight LLM runtime.

  2. Spin up the machines

   # Assume you have 3 Ubuntu 22.04 instances, SSH key-access set
   for i in 1 2 3; do
     cat <<EOF > /tmp/server$i.sh
     #!/bin/bash
     curl -sfL https://github.com/smitheryai/cord/releases/download/v1.0.0/cord-linux-amd64.gz | gunzip > /usr/local/bin/cord
     chmod +x /usr/local/bin/cord
     cord start --config /etc/cord.yml
     EOF
     scp /tmp/server$i.sh ubuntu@server$i:~/start.sh
     ssh ubuntu@server$i "bash ~/start.sh &"
   done
Enter fullscreen mode Exit fullscreen mode

Each machine now hosts a MCP server listening on :8000.

  1. Let Cord discover each other We’re using semantic discovery—Cord probes every reachable address and registers the available LLM or agent services automatically. 
   # /etc/cord.yml
   servers:
     - name: mcp-1
       address: 10.0.0.1:8000
     - name: mcp-2
       address: 10.0.0.2:8000
     - name: mcp-3
       address: 10.0.0.3:8000
   # No need to hard-code ports for agents
Enter fullscreen mode Exit fullscreen mode
  1. Configure the LLM CLI Use the same catalog entry on every node; Cord will auto-pick the nearest accelerator. 
   curl -sfL https://smithery.ai/servers/vishar-rumbling/install.sh | sh -s -- --llm=gpt-4
Enter fullscreen mode Exit fullscreen mode

The script writes a .cfg file pointing to the local MCP endpoint. It also registers a semantic profile named gpt-4-cli, so any agent can pull it from the mesh.

  1. Test the federation Deploy a thin “echo” agent on each node and ask them to call each other. 
   cord agent run --name echo --cmd "python - <<'PY'\nprint('Hello from', __file__)\nPY"
   # From node 1
   cord agent rpc echo@node2 "ping"   # Should return "pong" from node2
Enter fullscreen mode Exit fullscreen mode
  1. Scale Add more nodes by simply appending to /etc/cord.yml and restarting Cord. No new registry entries, no service discovery floods.

Why this matters

  • Zero-touch networking – Cord handles hops, NATs, and TLS without you writing any netcat hacks.
  • Semantic discovery – Agents describe the capabilities they expose (LLM, storage, compute) and query the mesh for the best match.
  • Dev-friendly – You can validate the entire stack locally in Docker, then spin up the same config on the cloud with terraform apply.

Next step: Check out the full catalog entry and copy the sample scripts. Plug the mesh into your own build pipeline, and let your agents go anywhere—not just where you’re staring. Happy hacking!

Get the catalog now and starter scripts.

Top comments (0)