PicoCluster Claw is a two-node appliance for running, testing, and benchmarking small AI models — a dedicated CUDA inference node paired with OpenClaw and ThreadWeaver. It's the reference platform for SMEEP — an open notebook of what small models can actually do in a real agent context.
14W idle. ~$2/month to keep running. Boots in seconds, stays up 24/7, and answers on your LAN — no API keys, no metered tokens, no traffic off the network.
SMEEP reference platform Dedicated CUDA node No API keys OpenClaw runtime Open source
Off-cluster inference, fragile orchestration, and shared workstation silicon all push private AI into the "later" pile. Claw collapses that work into one boxed appliance.
Frontier APIs run $15–60 per million tokens. A working agent burns millions a week, and every token you ship is data on someone else's storage policy.
Ollama, a web UI, a router, a fistful of scripts — any one update can break the pipeline. You end up maintaining plumbing instead of shipping work.
Once a request crosses your LAN, your code, designs, and customer data are governed by someone else's terms. Workstation-class hardware can't rewrite that contract.
Your laptop GPU is already pushing your IDE and browser. Phone-class SoCs thermal-limit under sustained load. You need dedicated execution silicon, not a background process fighting for cycles.
ClusterClaw hosts the OpenClaw automation runtime. ClusterCrush handles inference on dedicated CUDA silicon. Together they form an always-on AI execution node that lives on your desk and answers on your LAN.
Raspberry Pi 5 (8 GB) hosting OpenClaw — the local automation runtime. Handles planning, memory, tool dispatch, and MCP server coordination. Ships with 28 built-in tools across five servers: LEDs, system, clustercrush bridge, time, and files.
Jetson Orin Nano Super (8 GB) with 1024 CUDA cores serving OpenAI-compatible inference over the LAN. Reserved silicon — nothing else competes for the GPU, and your workstation stays free for its own work.
A Pimoroni Blinkt! strip driven by its own MCP server — an MCP for physical feedback. The runtime drives the strip to surface inference activity, tool calls, and alerts without a screen. The cluster's state is visible at a glance.
Five MCP servers expose the cluster as callable surface area: LED control, system metrics, the inference bridge to Crush, time and scheduling, and sandboxed file access. The runtime drives the box through its tools, not through config files.
ThreadWeaver is the chat front-end that ships with the appliance. It runs on the AgentStateGraph (ASG) substrate and can route any thread to on-cluster inference, an external provider, or both — selected per session, not system-wide.
On-cluster inference (Ollama), OpenAI, Anthropic, Mistral, and Groq all sit behind the same interface. Credentials live in one config, routing happens per-thread, and the active provider is always visible.
Threads are backed by the AgentStateGraph — not a flat JSON log. Conversations are stateful, resumable, and searchable without any extra setup.
Drop a Python function into the hook directory and it runs on every message — pre-send, post-receive, or both. ThreadWeaver's MCP + ASG + filter pipeline means you control what the model sees.
Thread history is encrypted on disk. On-cluster inference never leaves the appliance, and any external provider call is opt-in per thread — not a system-wide toggle.
Measured on the appliance, not marketing estimates.
| Component | Node | Detail |
|---|---|---|
| ClusterClaw | Raspberry Pi 5 | 8 GB LPDDR4X, BCM2712 quad-core A76 |
| ClusterCrush | Jetson Orin Nano Super | 8 GB unified memory, 1024-core Ampere GPU |
| Storage | Both nodes | NVMe SSD (size configurable) |
| Network | Both nodes | Gigabit Ethernet + Wi-Fi 5 |
| OS | ClusterClaw | Raspberry Pi OS (64-bit) |
| OS | ClusterCrush | JetPack 6 (Ubuntu 22.04 + CUDA) |
| LED strip | ClusterClaw | Pimoroni Blinkt! (8× APA102) |
The Claw appliance is the flagship deployment: plug in, power on, runtime up in minutes. The same OpenClaw + ThreadWeaver stack also installs on Mac, Linux, and Windows when you'd rather run it on the workstation you already have.
The two-node appliance. ClusterClaw hosts the OpenClaw automation runtime. ClusterCrush serves inference on dedicated CUDA silicon. Ships pre-built and pre-configured — no assembly, no provisioning.
Docker Compose overlay for Apple Silicon and Intel Macs. OpenClaw runtime, ThreadWeaver, and the Ollama bridge come up with one command — Metal GPU handles inference.
macOS install →Native install or Docker Compose. Full MCP tool surface, Ollama bridge, and operator console — identical behavior to the appliance.
Linux install →WSL2 + Docker Compose overlay. Bring the OpenClaw runtime and ThreadWeaver up from Windows Terminal — same stack, same tools, same behavior.
Windows install →git pull.
PicoCluster Claw is the reference platform for SMEEP — a living record of what small models can actually do in a real agent context. Every result is reproducible. Every script is open source.
# Clone and run on any Ollama host
git clone https://github.com/picocluster/PicoCluster-Claw
cd PicoCluster-Claw
# T1–T4 tool-call ladder
python3 scripts/bench-openclaw.py \
--models granite4.1:8b,nemotron-3-nano:4b
# 30-task capability suite
python3 scripts/bench-claw.py \
--models qwen2.5:7b,granite4.1:8b \
--output results/my-bench.jsonPicoCluster Claw ships pre-configured — OpenClaw automation runtime, dedicated inference node, ThreadWeaver console, 28 MCP tools. Plug in, power on, runtime up in minutes. No subscription, no external dependency, under $2/month to keep online.
Not ready for the appliance? Install the same stack on Mac, Linux, or Windows — your machine, identical behavior.