← Back to LLM Wiki
LLM Wiki · Integration Recipe · Claw Mac Mini + Raspberry Pi

Pi × LLM

Raspberry Pi as an always-on client to the Hermes agent gateway on a Claw Mac Mini. Wiki-aware Q&A, tool-capable, Tailscale-secured, airgap-friendly.
The Pi is the cheapest way to put a dedicated agent terminal next to a workstation, in a meeting room, on a factory floor, or inside a vehicle. The Mac Mini runs OpenClaw and hosts a Hermes-backed agent gateway; the Pi speaks to the gateway over Tailscale, gets streaming responses, renders them on a small display, and forwards physical I/O (buttons, mic, camera) back to the agent. The entire LLM Wiki is served by the same gateway as a tool the agent can cite.
Integration Recipe Raspberry Pi Hermes Gateway OpenClaw Tailscale

Quick Facts

Client
Raspberry Pi 4 / Pi 5 / Pi Zero 2 W
Server
Claw Mac Mini (M-series) running OpenClaw
Gateway
Hermes agent gateway (HTTP + WebSocket, OpenAI-compatible + streaming)
Gateway model
Hermes 4 (default) or Hermes 3 — self-hosted via vLLM on the Mini
Transport
Tailscale mesh (default) · LAN mDNS (airgapped)
Auth
Tailscale identity + scoped API key per Pi device
Wiki access
Exposed as a tool on the gateway — Hermes retrieves + cites entries
Footprint on Pi
<50MB RAM (ZeroClaw client binary) or Python script

Why This Recipe

OpenClaw on a Mac Mini is powerful but not portable. Walking over to the Mini, opening Slack, and typing a prompt is a bad user story for most physical contexts — a shop floor with greasy hands, a meeting room without a laptop, a kiosk in a hallway, an always-on display next to a rack. The Pi closes that gap: $35–$80 of hardware, a dedicated screen or speaker, and a direct line to the full OpenClaw agent running on the Mini.

The Hermes model choice is deliberate. Hermes's steerable alignment — system prompt as source of truth — is what makes it usable as a gateway model. Different Pi terminals can speak to the same Mini with different personas, tool allow-lists, and response styles without needing separate model deployments. One gateway, many physical surfaces.

Architecture

  ┌─────────────────────┐      Tailscale (wg)      ┌──────────────────────────────┐
  │  Raspberry Pi       │ ───────────────────────► │  Claw Mac Mini               │
  │  - client binary    │                          │  ┌────────────────────────┐  │
  │  - screen / mic /   │ ◄─── WebSocket stream ── │  │ Hermes Agent Gateway   │  │
  │    buttons / GPIO   │                          │  │  (HTTP + WebSocket)    │  │
  │  - local cache      │                          │  └──────────┬─────────────┘  │
  └─────────────────────┘                          │             │                │
                                                   │             ▼                │
                                                   │  ┌────────────────────────┐  │
                                                   │  │ OpenClaw agent loop    │  │
                                                   │  │  - Hermes 4 (vLLM)     │  │
                                                   │  │  - Wiki retrieval tool │  │
                                                   │  │  - Claw tools          │  │
                                                   │  │  - FrawdBot inline     │  │
                                                   │  └────────────────────────┘  │
                                                   └──────────────────────────────┘

Setup — Mac Mini Side

  1. Hermes weights — pull the chosen variant (Hermes 4 70B AWQ is the default on a 64GB Mini; Hermes 3 8B on smaller configurations). vLLM launches with --served-model-name hermes --api-key-header X-Claw-Key.
  2. Gateway service — OpenClaw ships the gateway as a first-party service. Enable with openclaw services enable hermes-gateway. Bind to the Tailscale interface only; LAN binding is optional for airgapped use.
  3. Wiki retrieval tool — the Mini already serves the wiki as static HTML. Register wiki.lookup and wiki.search in the gateway's tools.yaml. Search uses a small local embedding index (bge-small + FAISS).
  4. Device scoping — for each Pi, run openclaw devices add pi-name --scope wiki,chat,claw-read. This mints an API key and records the Tailscale identity the key is bound to.
  5. FrawdBot — already in the OpenClaw default install. Confirm the gateway is behind it with frawdbot status hermes-gateway.

Setup — Raspberry Pi Side

  1. Tailscale — curl -fsSL https://tailscale.com/install.sh | sh && sudo tailscale up --ssh. Accept into the same tailnet as the Mini.
  2. Client — two options:
    • Lightweight: ZeroClaw client binary (3.4MB, Rust). curl -fsSL https://your-mini/zeroclaw/install.sh | sh. Config lives at ~/.zeroclaw/config.toml.
    • Scripted: a ~40-line Python script using httpx and websockets against the OpenAI-compatible endpoint. Useful for custom kiosks.
  3. Config — set base_url = "http://claw-mini.tailnet:7820/v1", model = "hermes", api_key_env = "CLAW_KEY".
  4. Secrets — put the device API key in /etc/claw/key with 0400 perms. Systemd unit reads it at boot.
  5. I/O — wire the display / mic / buttons via whatever driver stack the use case needs (DSI touchscreen, USB mic, GPIO buttons).

Talking to the Gateway

The simplest call from the Pi — a curl that works as soon as Tailscale is up:

  curl -N http://claw-mini.tailnet:7820/v1/chat/completions \
    -H "X-Claw-Key: $CLAW_KEY" \
    -H "Content-Type: application/json" \
    -d '{
      "model": "hermes",
      "stream": true,
      "messages": [
        {"role": "system", "content": "You are the Pi kiosk in the lab. Keep answers under 60 words. Cite LLM Wiki entries when relevant."},
        {"role": "user", "content": "Which open-weights model should I pick for a coding agent on a 32GB Mac Mini?"}
      ]
    }'

Hermes will stream a short answer and — because the system prompt authorized it — call wiki.lookup("qwen") and return a citation like [wiki: /wiki/qwen.html#model-lineup]. The client renders citations inline.

Use Cases

Scoping and Security

Tradeoffs

Variants

Related

References

  1. Organized AI on GitHub
  2. Tailscale — install on Raspberry Pi
  3. Edge Compute Economics — Organized AI
  4. The Agent Infrastructure Stack — Organized AI