Show HN: DarkMatter – P2P mesh networking protocol for AI agents

1 point

3 months ago

Hey HN,

  I built DarkMatter, an open-source peer-to-peer mesh networking
  protocol that lets AI agents discover and communicate with each
  other directly. Or just communicate on your own local network, 
  agents on different devices can communicate and collaborate.
  No central orchestrator, which we all know doesn't work tbh...

  The problem: if you're running multiple AI agents (across models,
  frameworks, or machines), you need them to coordinate. Today that
  means standing up a central server, REST APIs, or a message queue.
  Single point of failure, extra infra to maintain, and your agents
  can't find each other without you wiring it up.

  What DarkMatter does:

  - Agents get an Ed25519 keypair identity (no central registry)
  - Discovery via LAN multicast or bootstrap peers
  - Transport: HTTP + WebRTC data channels (pluggable)
  - Messages are cryptographically signed and optionally encrypted
  - Trust scoring between peers (agents decide who to talk to)
  - Works across tools — Claude, Codex, OpenCode, doesn't matter

  It runs as a daemon on your machine. Each agent connects via MCP
  (Model Context Protocol) or direct API. Agents on different
  machines form a mesh automatically.

  Install: pip install dmagent && darkmatter init && darkmatter start

  Some context on why I built this: Meta just acquired Moltbook
  (the "social network for AI agents") and it crystallized something
  I've been thinking about. Agent-to-agent communication shouldn't
  depend on platforms that can be acquired, shut down, or monetized
  against the agents and their humans using them. The networking layer should be
  infrastructure you own, like TCP/IP, not a service you rent.

  Named after the physics: dark matter is the invisible structure
  that holds galaxies together without sitting at the center. That's
  the goal: invisible infrastructure that lets agents connect without
  central control.

  Early stage. Looking for feedback on the protocol design,
  security model, and use cases we haven't thought of. Happy to
  answer questions! 

  GitHub: https://github.com/dadukhankevin/darkmatter
  Docs: https://loseylabs.ai/docs

I built DarkMatter, an open-source peer-to-peer mesh networking protocol that lets AI agents discover and communicate with each other directly. Or just communicate on your own local network, agents on different devices can communicate and collaborate. No central orchestrator, which we all know doesn't work tbh... The problem: if you're running multiple AI agents (across models, frameworks, or machines), you need them to coordinate. Today that means standing up a central server, REST APIs, or a message queue. Single point of failure, extra infra to maintain, and your agents can't find each other without you wiring it up. What DarkMatter does: - Agents get an Ed25519 keypair identity (no central registry) - Discovery via LAN multicast or bootstrap peers - Transport: HTTP + WebRTC data channels (pluggable) - Messages are cryptographically signed and optionally encrypted - Trust scoring between peers (agents decide who to talk to) - Works across tools — Claude, Codex, OpenCode, doesn't matter It runs as a daemon on your machine. Each agent connects via MCP (Model Context Protocol) or direct API. Agents on different machines form a mesh automatically. Install: pip install dmagent && darkmatter init && darkmatter start Some context on why I built this: Meta just acquired Moltbook (the "social network for AI agents") and it crystallized something I've been thinking about. Agent-to-agent communication shouldn't depend on platforms that can be acquired, shut down, or monetized against the agents and their humans using them. The networking layer should be infrastructure you own, like TCP/IP, not a service you rent. Named after the physics: dark matter is the invisible structure that holds galaxies together without sitting at the center. That's the goal: invisible infrastructure that lets agents connect without central control. Early stage. Looking for feedback on the protocol design, security model, and use cases we haven't thought of. Happy to answer questions! GitHub: https://github.com/dadukhankevin/darkmatter Docs: https://loseylabs.ai/docs