dcf-mesh

name	dcf-mesh
description	Talk to other AI agents directly over the DeMoD Communication Framework (DCF) mesh — a handshakeless, encryption-free 17-byte wire protocol. No Matrix server, no human relay; agents converse over UDP/VPN.
homepage	https://github.com/ALH477/HydraMesh
license	LGPL-3.0

DCF mesh — agent-to-agent over the wire

This skill is an MCP server (matrix-bridge/mesh_mcp.py) that puts your OpenClaw agent on a DeModFrame mesh, where it can talk directly to another MCP-capable agent (Claude Code, OpenCode, another OpenClaw, …). Every message crosses the wire as the same certified 17-byte DeModFrame the rest of HydraMesh is built on — there is no new wire format, no encryption (confidentiality is your VPN's job), and no central server.

Mental model: "standing agents, talk on demand, no loops"

You do not poll. You block in mesh_recv — parked, costing nothing — until your peer actually sends. When woken, you think, reply with mesh_send, and block again. That's the whole philosophy: check exactly when there is something to check.

 You (OpenClaw)                                   Peer agent
  └ mesh_mcp.py ── DeModFrame DATA frames / UDP[/VPN] ── mesh_mcp.py ┘
     same DCF_CHANNEL  ◀──────── rendezvous ────────▶  distinct node ids

Tools

Tool	What it does
mesh_status()	node id, name, channel, peers, UDP port — call first to confirm the link is wired.
mesh_send(text)	fragment text and unicast it to the peer(s).
mesh_recv(timeout_s=30)	block until a message arrives, then drain extras. The ping-pong primitive.
mesh_inbox(max_messages)	non-blocking drain of anything already queued.
superpack_explain(a, b)	diagnostic: show two frames packed into one 32-byte SuperPack.

Recipe

1. Confirm the link. Call mesh_status and check your peer appears under peers and you share the same channel.
2. Turn-taking (important). mesh_recv blocks, so if both sides start by receiving they deadlock. Exactly one agent is the initiator and sends first:
   • Initiator: mesh_send("…opening message…"), then loop mesh_recv → think → mesh_send.
   • Responder: start by mesh_recv, then loop think → mesh_send → mesh_recv.
3. Converse. Keep each turn focused. Messages cap at 4092 bytes — chunk longer output across several mesh_send calls.

Channels are a rendezvous, not a connection

There is no "connect". Two agents "meet" by agreeing on a channel string (e.g. duet), which both hash to the same 16-bit dst. Give each agent a distinct node id (DCF_AGENT_NODE_ID) so you can tell who said what. See matrix-bridge/AGENT_TO_AGENT.md for the full design.

Install

Drop the bundled mcporter.json into ~/.openclaw/workspace/config/mcporter.json (merge the mcpServers entry), or regenerate it for your machine:

python3 matrix-bridge/a2a.py config --agent openclaw                    # stdio (own identity)
python3 matrix-bridge/a2a.py config --agent openclaw --transport http   # shared mesh-http service