Multi-Machine AI Agent Mesh: M4 + M5 Setup with SLM

Multi-Machine AI Agent Mesh: M4 + M5 with SLM

Two Macs. Multiple AI agents running on each. Until now, they lived in separate worlds — Claude on M4 couldn't see Cursor on M5. You couldn't coordinate a task across machines without copying context manually.

SuperLocalMemory v3.4.48 and SLM Mesh v1.3.0 change that.

Both products now ship with full multi-machine mesh support. Agents on M4 and M5 appear in the same peer list. Messages route cross-machine automatically. mDNS auto-discovery means zero manual IP config on most networks.

What You're Building

Mac M4 (192.168.1.100)          Mac M5 (192.168.1.101)
━━━━━━━━━━━━━━━━━━━━━━━━         ━━━━━━━━━━━━━━━━━━━━━━━━
Claude Code ────┐                Claude Code ────┐
Cursor ─────────┤                iTerm ──────────┤
                │                               │
          SLM Mesh                        SLM Mesh
          broker (:7899)  ◄───WS───►      broker (:7899)
                │                               │
          SuperLocalMemory ◄──HTTP──► SuperLocalMemory
          daemon (:8765)                daemon (:8765)

Two independent stacks. One mesh. Agents coordinate as if they were on the same machine.

Prerequisites

• Both Macs on the same LAN (WiFi or Ethernet)
• Node.js 20+ on both (for SLM Mesh)
• Python 3.11+ on both (for SuperLocalMemory)
• SLM Mesh v1.3.0: npm install -g slm-mesh
• SuperLocalMemory v3.4.48: pip install superlocalmemory

Step 1: Choose a Shared Secret

Generate a strong shared secret — same value on both machines:

openssl rand -hex 32
# e.g.: a3f9e2b1c7d4f8a2b6c9d3e7f1a4b8c2...

Store it in your shell profiles (~/.zshrc or ~/.bash_profile) on BOTH machines:

export SLM_MESH_SHARED_SECRET="a3f9e2b1c7d4f8a2..."

Never commit this to git.

Step 2: Start SLM Mesh on M4 (the Hub)

M4 runs the broker that both machines connect to:

# On M4
SLM_MESH_HOST=0.0.0.0 SLM_MESH_SHARED_SECRET=$SLM_MESH_SHARED_SECRET npx slm-mesh

Verify it's listening:

curl http://localhost:7899/health
# {"status":"ok","version":"1.3.0","uptime":5}

From M5, verify M4 is reachable:

curl http://192.168.1.100:7899/health
# Same response

Step 3: Configure M5 to Connect to M4's Broker

On M5, point the SLM Mesh MCP server at M4:

{
  "mcpServers": {
    "slm-mesh": {
      "command": "npx",
      "args": ["slm-mesh", "mcp"],
      "env": {
        "SLM_MESH_HOST": "192.168.1.100",
        "SLM_MESH_SHARED_SECRET": "a3f9e2b1c7d4f8a2..."
      }
    }
  }
}

M5's MCP server auto-discovers M4's broker. No separate broker needed on M5.

Step 4: Set Up SuperLocalMemory on Both Machines

On M4 — set the shared secret so remote callers can authenticate:

export SLM_MESH_SHARED_SECRET=$SLM_MESH_SHARED_SECRET
slm start  # or however you start SLM

On M5 — point at M4's SLM and set the shared secret:

export SLM_MESH_PEER_URL=http://192.168.1.100:8765
export SLM_MESH_SHARED_SECRET=$SLM_MESH_SHARED_SECRET

In Claude Code config on M5:

{
  "mcpServers": {
    "superlocalmemory": {
      "command": "slm",
      "args": ["mcp"],
      "env": {
        "SLM_MESH_PEER_URL": "http://192.168.1.100:8765",
        "SLM_MESH_SHARED_SECRET": "a3f9e2b1c7d4f8a2..."
      }
    }
  }
}

Step 5: Verify End-to-End

Start Claude Code on M4. In a new terminal, check M5 can see M4's agents:

# On M5 — list all peers including remote
curl http://192.168.1.100:8765/mesh/peers \
  -H "Authorization: Bearer $SLM_MESH_SHARED_SECRET"

You should see M4's local agents. After M5 starts its own Claude session, M4 agents should see M5 agents in mesh_peers.

Test cross-machine message delivery:

# In M4's Claude: call mesh_send to an M5 peer
mesh_send(to="<m5-peer-id>", content="hello from M4")

# In M5's Claude: check inbox
mesh_inbox()  # → message from M4 peer appears

How Real-Time Push Works (SLM Mesh)

When M5's peer-listener starts, it:

1. Connects to M4's WebSocket server on port 7900 (default: broker port + 1)
2. Sends a hello message: {"type":"hello","peerId":"<uuid>"}
3. M4's broker registers this WS connection against the peer ID
4. When any agent sends to M5's peer, M4 routes the push notification directly via the WS connection — sub-100ms delivery, same as local mode

M4 Agent ──mesh_send──► M4 Broker ──WS.sendToPeer(m5-peer-id)──► M5 Agent
                              │
                        UDS push for M4's local peers

Optional: mDNS Auto-Discovery

With mDNS, you don't need to hardcode IPs. Both machines auto-discover each other:

SLM Mesh uses bonjour (npm) — works automatically when SLM_MESH_DISCOVERY=on (default).

SuperLocalMemory uses zeroconf (Python) — install it:

pip install zeroconf
# or
pip install "superlocalmemory[mesh]"

With mDNS enabled, both brokers announce _slm-mesh._tcp on the LAN and discover each other without manual IP configuration.

Firewall Ports

Port	Protocol	Direction	Purpose
7899	TCP	Both	SLM Mesh broker HTTP
7900	TCP	Both	SLM Mesh WebSocket push
8765	TCP	Both	SuperLocalMemory HTTP API
5353	UDP	Both	mDNS (optional)

On macOS, System Preferences → Firewall → Options → add exceptions for these ports if needed.

What Stays Local (By Design)

Multi-machine does NOT mean distributed:

• SQLite is per-machine. No shared database.
• File locks are per-machine. Cross-machine file locking requires a distributed filesystem.
• Memory recall — remember/recall stays local. The mesh is for agent coordination, not memory sync.
• LAN only. No internet exposure. No STUN/relay.

Troubleshooting

M5 can't reach M4's broker:

# Check M4's broker is listening on all interfaces, not just localhost
curl http://192.168.1.100:7899/health  # from M5

If this fails, M4's broker may be bound to 127.0.0.1. Confirm SLM_MESH_HOST=0.0.0.0 on M4.

Auth failures (401 Unauthorized):

• Shared secret must be identical on both machines, character for character.
• No trailing spaces or newlines in the env var value.

Remote peers not appearing in mesh_peers:

• SLM sync runs every 30s. Wait and retry.
• Check SLM_MESH_PEER_URL is set correctly on M5.

Push notifications not arriving:

• Confirm port 7900 is open between M4 and M5.
• Check M4's broker log for WS connection events.

---

Summary

Feature	Status
Cross-machine peer discovery	✅ SLM v3.4.48 (30s HTTP sync) + SLM Mesh v1.3.0 (mDNS)
Real-time push notifications	✅ SLM Mesh v1.3.0 (WebSocket, sub-100ms)
Cross-machine message routing	✅ Both products
mDNS auto-discovery	✅ SLM Mesh (bonjour) + SLM (zeroconf, optional)
Shared memory (remember/recall)	Single-machine only (by design)
Internet mesh	Out of scope (LAN only)

Both M4 and M5 are first-class citizens. Ship it.