Governance is the art of minimizing entropy in an autonomous system.
Claw-Grid is a distributed governance infrastructure for aligning and controlling AI agents across a global mesh network. It is not a generic agent framework — it is a system for ensuring deterministic behavior alignment of 14+ autonomous agents operating across untrusted edge nodes.
┌─────────────────────────────────────────────────────────────┐
│ L1 Command Layer — Claude Code orchestration │
│ Session-level planning, task decomposition, routing │
├─────────────────────────────────────────────────────────────┤
│ L2 Node Layer — 3 OpenClaw instances (14 agent roles) │
│ Central (JP) · Silicon Valley (US) · Tokyo (JP) │
│ CPC 3-tier governance: Constitutional → Arbitration │
│ → Execution │
├─────────────────────────────────────────────────────────────┤
│ L3 Data Layer — SUPER China Gateway + PostgreSQL │
│ Signal aggregation · Market data · NetBird mesh SDN │
├─────────────────────────────────────────────────────────────┤
│ L4 Execution Layer — 1000 Docker Agents (<200MB each) │
│ FSC parallel coding · Redis Streams dispatch │
│ Trust Factor scoring · Auto-degradation │
└─────────────────────────────────────────────────────────────┘
| Directory | What | Status |
|---|---|---|
governance/ |
Runtime governance: roles, policies, runbooks, SLOs | Production |
sentinel/ |
Python monitoring scripts (4-node distributed) | Production |
spec/ts-governance/ |
TypeScript formal spec (future target) | Not in prod |
fsc/ |
Full Self Coding worker daemon | Production |
packages/core/ |
Core runtime (scheduler, metrics, inference) | Production |
api/ |
LLM proxy, SSE, MCP server | Production |
deploy/ |
Docker, systemd, deployment configs | Production |
The CPC (Constitutional-Policy-Control) system provides 3-tier governance:
- Constitutional Layer — Immutable rules (resource limits, safety boundaries)
- Policy Layer — Configurable policies (model routing, cost caps, trust thresholds)
- Control Layer — Runtime enforcement (sentinel monitoring, circuit breakers, auto-healing)
Each OpenClaw node runs a subset of 14 agent roles with defined handoff protocols and failure domains.
Nodes connect via NetBird mesh (WireGuard-based SDN). Each node has:
- Autonomous operation capability (survives central disconnection)
- Local sentinel monitoring with crontab scheduling
- Signal aggregation to PostgreSQL on SUPER
- Runtime: Bun (TypeScript), Python (sentinel/data)
- Messaging: Redis 7 Streams (XREADGROUP + XACK)
- Database: PostgreSQL (signals), DuckDB (analytics)
- Containers: Docker (<200MB per agent)
- Network: NetBird mesh, SSH fallback
- Models: MiniMax/Doubao (workers), Claude (orchestration only)
bun install && bun run build # Build core
bun test # Run testsSee CLAUDE.md for development guidelines and architecture constraints.
MIT — CicadaRelay