All projects
Live · My own product

Braincore Pro

Decentralized AI agent swarm

The problem

Central orchestration is a single point of failure: if the brain dies, all agents stop. I wanted a cluster where any node can die without stopping the work, and where state converges without a central database.

The solution

Braincore Pro is a self-hosted agent platform. There is no central orchestrator: every node announces itself, membership converges through epidemy-style gossip, and tasks are published to all peers. Whichever node is free claims the task and the claim is resolved as a last-write-wins CRDT write, so exactly one executor wins under simultaneous claims.

The coordination layer is stigmergic: agents leave pheromone trails with a 30-second TTL and true exponential decay, and the strongest trail gets the work. Live state is exposed over a WebSocket feed, so you can watch nodes, the shared board and pheromone decay in real time.

It runs on Node.js built-ins only — no npm dependencies in the runtime, which makes it portable and easy to audit.

Verifiable proof

Each item is checkable — a command, a public URL, or a number measured on the live system.

  • 3 nodes in the running cluster, 2 peers alive
  • ~210,000 gossip messages sent in the current run, 0 rejected
  • Test suite: 281 tests, 277 passing — the 3 known failures are named openly below and are being fixed
  • Verified myself: WebSocket /events answers 101 Switching Protocols and pushes ~1 snapshot/s
  • Submitted one task through the machine gate: 2 members confirmed → CRDT (3 entries) → 2 pheromones → executed ok:true → pheromones decayed to 0
  • Zero runtime dependencies: repo package.json has empty dependencies, and one test asserts only Node built-ins are used
  • Public test output in the repository (t4-1997395602.txt) shows node discovery, CRDT anti-double-claim, CRDT convergence, pheromone decay and the hand-written RESP client passing
  • Known failures (honest): trace spans for tool calls, automatic episodic memory recording, and the handoff chain test — instrumentation of tool calls, being fixed