Inter-gateway coordination layer for OpenClaw clusters.
Hive connects independent OpenClaw gateway instances via MQTT, enabling cross-gateway commands, event-driven file sync, sentinel monitoring, and machine-to-machine heartbeats — all without burning LLM tokens on deterministic operations.
Running multiple OpenClaw gateways creates coordination problems:
- Heartbeat tax — each gateway wakes its LLM every 30 minutes to check if anything happened. 90%+ of the time, nothing did. That's thousands of wasted tokens per day.
- File divergence — boot files, memory files, and configs drift across instances without a sync mechanism.
- No cross-gateway communication — gateways can't delegate work to each other.
- Redundant monitoring — every gateway independently checks the same things (email, calendar, services).
Hive solves these by putting a thin daemon on each box that handles coordination deterministically, and only wakes the LLM when something actually needs judgment.
MQTT (hive topics)
↕
[hive-daemon] ← thin sidecar, one per box
↕ ↕
system event hive-cli
↕ ↕
[OpenClaw instance(s)]
↕
[hive bridge plugin] ← in-process runtime event tap (optional)
- Inbound to OC: daemon →
openclaw system event→ LLM wakes with specific context - Outbound from OC: LLM →
hive-cli send→ daemon → MQTT - Deterministic work: daemon →
hive-daemon.d/handler scripts → no LLM involved - Real-time runtime telemetry (optional): OpenClaw plugin observes runtime/observability events and publishes session/event deltas to Hive topics
| Component | Purpose |
|---|---|
daemon/ |
Always-running sidecar — MQTT routing, heartbeats, handler dispatch |
cli/ |
Stateless CLI — send commands, check status, reply to requests |
plugin/ |
OpenClaw extension package(s) — in-process bridge for runtime observability and low-latency presence/event publication |
skills/ |
OpenClaw skills — teach the LLM how to participate in the hive |
contrib/handlers/ |
Example hive-daemon.d/ scripts (git-sync, health-check, etc.) |
contrib/shared-memory/ |
Git-based cluster memory pattern for multi-gateway deployments |
Instead of waking an LLM every 30 minutes to ask "anything happening?", hive uses zero-cost daemon scripts that monitor continuously and only wake the LLM when something needs a brain:
Before: every 30 min → wake LLM (18k tokens) → "anything?" → "no" → waste
After: every 1-5 min → daemon script (0 tokens) → "no" → free
...until something actually happens → wake LLM with context
See docs/protocol.md for the full design.
Coming soon — Phase 1 implementation in progress.
MIT