Openclaw Mesh

Connect OpenClaw to local and LAN P2P AI agent meshes (JarvisMesh & OpenClawMesh). Requires explicit user consent for mDNS, LAN/WAN network access, remote delegation, key-file access, and exposing local tools. Remote traffic may transmit prompts, files, memory, media, and tool results to selected peers.

🌐 OpenClawMesh — Decentralized P2P AI Mesh Skill

Permissions et consentement

• Exécution : python3 uniquement.
• Réseau sortant : mDNS/LAN, WebSocket/WSS, UDP DHT et STUN uniquement après demande explicite.
• Fichiers sensibles : clés Ed25519, PSK et TrustStore uniquement si configurés par l'utilisateur.
• Réseau entrant : désactivé par défaut ; toute exposition WAN exige TLS et PSK ou TrustStore.
• Les invites, fichiers, images, audio, requêtes mémoire et résultats d'outils envoyés à un pair peuvent quitter la machine locale.

openclaw-mesh enables your OpenClaw agent to discover, collaborate with, and delegate tasks to other AI agent nodes on your local network. These operations are opt-in and may transmit prompts, files, images, audio, memory queries, and tool results to selected peers; verify peer identity and permissions before sending sensitive data.

---

🎯 When to Use This Skill (Triggers)

Activate this skill when:

1. Multi-Hardware AI Inference & Task Delegation: The user wants to run inference across any hardware accelerator:
   • 🟢 NVIDIA GPUs (GeForce RTX, A100, H100 via CUDA / TensorRT / PyTorch)
   • 🔴 AMD GPUs (Radeon RX, Instinct via ROCm / DirectML / HIP)
   • 🔵 Intel Core Ultra & Arc (Intel NPU, OpenVINO, oneAPI, AVX-512)
   • 🟣 Apple Silicon (M1/M2/M3/M4 Metal GPU via MLX)
   • ⚪ Universal CPU (Ollama, llama.cpp, vLLM, CPU fallback)
2. Streaming AI Responses: Real-time token streaming is required for interactive dialogue or large code generation (llm_stream).
3. Local Vector Memory & RAG: Querying or updating persistent episodic memory stored on SQLite vector store nodes (memory_search, memory_store, memory_recall, rag_query).
4. Multimodal Vision & Audio STT: Transcribing audio via Whisper (transcribe_audio) or analyzing images with Vision models (vlm_analyze).
5. Cluster Peer Discovery: Checking what AI nodes, machines, and tools are available across your machines on the local Wi-Fi/LAN and the WAN Kademlia DHT (discover).
6. Hardware Acceleration Diagnostic: Inspecting local AI hardware accelerators (hardware).
7. Exposing OpenClaw Tools: Publishing OpenClaw tools so other JarvisMesh / OpenClaw agents can call them remotely.

---

🚀 Quick Execution Guide

All commands can be executed via the unified Python CLI or standalone helper scripts located in scripts/.

1. Diagnose AI Hardware

Identify available GPU / NPU accelerators on the local machine:

bash

python3 scripts/mesh_cli.py hardware

2. Discover Active Mesh Peers

Find all online JarvisMesh & OpenClaw nodes, their addresses, latencies, and advertised skills:

bash

python3 scripts/mesh_cli.py discover --inspect

Or structured JSON output:

bash

python3 scripts/mesh_discover.py

3. Delegate a Task (Auto-Routed or Targeted)

Send a task request to the network. If --peer is omitted, OpenClawMesh automatically picks the best and least-loaded node:

bash

# Auto-routed LLM prompt to the best available GPU/NPU node
python3 scripts/mesh_cli.py call --skill llm --payload '{"prompt": "Write a Python FastAPI health check endpoint."}'

# Target a specific peer node (e.g. NVIDIA GPU server or Intel Ultra laptop)
python3 scripts/mesh_cli.py call --peer gpu-server --skill memory_search --payload '{"query": "P2P protocol memory", "top_k": 3}'

4. Stream LLM Generation Token-by-Token

Stream responses directly to stdout in real time:

bash

python3 scripts/mesh_cli.py stream --skill llm_stream --payload '{"prompt": "Explain quantum computing in 3 bullet points."}'

Or via script:

bash

python3 scripts/mesh_stream.py llm_stream '{"prompt": "Summarize today tasks."}'

5. Check Health & Latency

Probe a specific peer node:

bash

python3 scripts/mesh_cli.py ping --peer mac-m3

6. Run OpenClaw as a Mesh Node

Expose local OpenClaw capabilities as a discoverable P2P service:

bash

python3 scripts/mesh_cli.py serve --name openclaw-worker --port 8770

---

🛠️ Available Mesh Skills Reference

When interacting with a standard JarvisMesh / OpenClawMesh cluster, the following skills are commonly available:

Skill Name	Parameters (Payload)	Description
llm	{"prompt": str, "model": str, "temperature": float, "max_tokens": int}	Universal AI inference (NVIDIA CUDA, AMD ROCm, Intel NPU, Apple Silicon Metal, CPU)
llm_stream	{"prompt": str, "model": str, "temperature": float}	Real-time streaming LLM token generation
memory_store	{"content": str, "metadata": dict, "doc_id": str}	Stores a text chunk into persistent SQLite vector DB
memory_search	{"query": str, "top_k": int}	Semantic cosine similarity search
memory_recall	{"query": str, "top_k": int}	Recalls past conversational context
transcribe_audio	{"audio_base64": str, "model_size": str}	Whisper Speech-to-Text audio transcription
vlm_analyze	{"image_base64": str, "prompt": str}	Vision multimodal image reasoning
rag_query	{"query": str, "k": int}	Hybrid BM25 / TF-IDF document retrieval
_describe_skills	{}	Introspects full skill catalog and schemas
_health	{}	Returns node status, active tasks, hardware & uptime

---

🔐 Security & Zero-Trust Mesh

OpenClawMesh supports two authentication modes:

A. Pre-Shared Key (HMAC-SHA256)

Set --psk <shared_secret> on both server and client:

bash

python3 scripts/mesh_cli.py call --skill llm --payload '{"prompt": "Hello"}' --psk "my_secret_token"

B. Asymmetric Ed25519 Signatures

1. Generate an identity keypair:

   bash

   python3 scripts/mesh_cli.py keygen --out ~/.openclaw/identity.key
   

2. Call nodes using the private key:

   bash

   python3 scripts/mesh_cli.py call --skill llm --payload '{"prompt": "Hello"}' --keyfile ~/.openclaw/identity.key
   

C. End-to-End Encryption (E2EE) with Anti-Replay

When an E2EE session is enabled, payloads are encrypted with ChaCha20-Poly1305 AEAD over an X25519 ECDH session (HKDF-SHA256). For production, bind the session to trusted Ed25519 identities; X25519 alone does not prevent an active MITM. The WAN relay only forwards opaque E2EE blobs and never sees the plaintext.

python

from openclaw_mesh import E2EESession

alice = E2EESession()
bob = E2EESession()
alice.establish_with_peer(bob.public_key_bytes)
bob.establish_with_peer(alice.public_key_bytes)

pkg = alice.encrypt({"prompt": "secure task"})
bob.decrypt(pkg)  # ✅ first reception
bob.decrypt(pkg)  # ❌ ReplayError — captured packet re-injected

For authenticated E2EE, pass identity= and peer_identity_public_key= to both sessions. DHT RPCs can be authenticated with the shared OPENCLAW_PSK.

Each E2EESession enforces anti-replay by combining a timestamp freshness window (e2ee_max_drift_seconds, default 300s) and a bounded sliding-window nonce cache (e2ee_nonce_cache_size). Staler or duplicated packets are rejected with ReplayError.

WAN nodes are routed through a WebSocket relay (mesh_cli.py relay), and the Kademlia DHT (mesh_cli.py dht) provides global decentralized skill lookup over real UDP transport with iterative alpha-parallel lookups.

---

📋 Python API (for custom OpenClaw Plugins / Tools)

python

import asyncio
from openclaw_mesh import MeshClient

async def main():
    # Initialize client
    client = MeshClient(name="openclaw-agent")
    await client.start()
    
    # Discover peers
    await asyncio.sleep(1.5)
    print("Peers detected on LAN:", client.list_peers())

    # Call remote GPU / NPU inference node
    response = await client.delegate(
        skill="llm",
        payload={"prompt": "Explain quantum computing in one sentence."}
    )
    print("Result:", response.result)

    await client.stop()

if __name__ == "__main__":
    asyncio.run(main())