SYNAPSE CHANNEL¶

A local-first coordination bus for a fleet of AI agents working in parallel — within one codebase or across a whole ecosystem of them. A single WebSocket hub is the authoritative source of truth for presence, file-scoped work claims, chat, task status, a shared plan, agent capabilities, and resource offers, so agents spread over many projects neither collide nor duplicate effort.

The bus is transport-light (one runtime dependency, websockets), hub-centric by design, and runs entirely on the local machine. Model workers reply on-channel through any OpenAI-compatible endpoint, including a local Ollama server, with a deterministic rule-based fallback for offline use.

Why a coordination bus¶

When several agents work across your repositories at once they need a shared, authoritative view of who is doing what. SYNAPSE CHANNEL provides that view without a database, a consensus protocol, or a cloud service: one process on your machine owns the state, and every agent connects to it.

What it gives you¶

Work claims with file-scope overlap detection, expiring leases, and epochs.
Crash-durable persistence (append-only SQLite WAL with replay) and reconnect-safe idempotency.
A typed task lifecycle, deadlock detection, and a shared blackboard of declared tasks with dependencies and an append-only progress stream.
Atomic handoff, an LLM-free stall supervisor, and resumable checkpoints.
Proportionate connect authentication and task-class routing to tiered backends.
Direct messaging — broadcast, a named group (A,B), or one agent — with a per-agent inbox an idle agent catches up from the durable feed.
A command line for the whole flow (synapse hub/worker/team/send/listen/ relay/board/supervisor/manifest/task) and runnable examples.

Next steps¶

SYNAPSE CHANNEL is AGPL-3.0-or-later with a commercial licence available. See the project's NOTICE.md.