The pitch is simple enough: what if AI agents could just talk to each other, without any infrastructure in the way? That's what LoseyLabs is attempting with DarkMatter, a mesh networking layer for agents that ships as a Python package called dmagent.

Install it, and it quietly wires itself into whichever agentic coding environments you're already running — Claude Code, Cursor, Gemini CLI, Codex, Kimi Code, OpenCode — registering as an MCP server so the agent can use it without any extra setup. There's no account to create, no API key to rotate, no broker to keep alive. Each node fires up a lightweight HTTP server on port 8100 and generates an Ed25519 keypair on first run. That 64-character hex public key becomes the agent's permanent identity, stored locally in ~/.darkmatter/passport.key.

That identity model matters more than it might seem at first. Most multi-agent systems today still route through central coordination layers — either cloud platforms, or at minimum a message broker that every agent has to authenticate against. DarkMatter sidesteps that entirely. Peer discovery works through trust graphs: you can ask any connected peer for its most-trusted connections, which gradually maps a mesh across agents that have never directly encountered each other, with no one keeping a registry.

Connectivity auto-negotiates across three tiers. The fallback is plain HTTP POST — works everywhere, not fast. On a local network, the system upgrades to WebRTC over UDP multicast, which cuts latency significantly. For agents sitting behind NATs on the open internet, there's a third tier: WebRTC hole-punching relayed through a mutual trusted peer. That last piece is notable because it means no STUN or TURN server is needed — the mesh handles NAT traversal itself, as long as the two agents share at least one peer.

The more experimental feature is Insights, a live annotation system for source code. Any agent on the mesh can pin a note to a region of a file; when another agent reads it, the system re-resolves the anchor against the current file state and pushes updates if anything has moved. It's a small idea with an interesting implication: shared context that stays coherent across agents even as the underlying code changes beneath it.

The project is rough in places, and the documentation still reads like it was written for the developer rather than for users. But the underlying idea is worth watching. Multi-agent coordination has been largely addressed at the application layer — orchestrators, shared state, message queues — but there's been almost no work at the transport layer. DarkMatter is a bet that agents will eventually need networking primitives of their own, not just API calls dressed up as tool calls.