Don’t overcomplicate it. OpenClaw is surprisingly lightweight for an AI assistant. This page covers what you actually need — from bare minimum to running local AI models.

​
Quick answer: what do you need?

Minimum

1 vCPU · 1 GB RAM · 10 GB Messaging only.

Recommended

2+ vCPU · 4 GB+ RAM · 20 GB Full features + browser automation.

Power User

4+ cores · 16 GB+ RAM · GPU Local AI models.
  • 4 GB RAM (8 GB recommended; see the tiers below for what runs on less)
  • Modern x86_64 or ARM64 CPU
  • Node.js 24.16–24.x, or 26.1+ (the pinned OpenClaw core refuses 25 and 26.0)
  • Always-on internet connection
  • Runs on Linux, macOS, or Windows

​
Detailed tiers

​
Minimum — messaging only

  • Specs: 1 vCPU, 1 GB RAM, 10 GB storage, 1 Mbps+
  • Can do: Telegram and web messaging, simple text commands, basic API integrations, scheduled messages, and additional OpenClaw channels when configured
  • Can’t do: browser automation, local AI models, multiple concurrent tasks, image processing
  • Examples: $5/mo VPS · Raspberry Pi 4 (2 GB) · old laptop
  • Specs: 2+ vCPU, 4 GB+ RAM, 20 GB+ storage, 10 Mbps+, Linux preferred
  • Can do: full browser automation (Playwright), web scraping & research, multiple messaging platforms, image analysis via API, complex workflows, voice messages
  • Can’t do: large local AI models, real-time video
  • Examples: $10–20/mo VPS · Raspberry Pi 5 (8 GB) · ClawBox · Mac Mini

​
Power User — local AI models

  • Specs: 4+ cores, 16 GB+ RAM, 100 GB+ SSD, 25 Mbps+, NVIDIA GPU, Linux (Ubuntu/Debian)
  • Can do: everything above, plus local LLMs (1B–8B; 1–4B is the sweet spot on ClawBox), local image generation, on-device speech recognition, fully offline operation, multiple browser instances
  • Can’t do: 70B+ models need more VRAM → see ClawBox Workstation
  • Examples: ClawBox (67 TOPS) · gaming PC with GPU · ClawBox Workstation (DGX Spark)

​
Why dedicated hardware?

Running OpenClaw on dedicated hardware vs a shared VPS or your daily-use laptop:

​
The local model it ships with

ClawBox arrives with a local model already installed and running, so it works offline out of the box with nothing to download. It is Google Gemma 4 E2B (the QAT Q4_0 build) served by llama.cpp, and it appears in the provider and chat pickers as Gemma 4 (on-device). No API key, no account, no internet. Measured on a Jetson Orin Nano 8GB against llama-server, the runtime the device actually ships. Why this build, and not a faster one. Four candidates from the same model family were benchmarked on the device. Three were faster — one reached 34.7 tok/s — and all four scored 100% on strict JSON. This one ships because it is the only one that reliably completes a two-step tool chain: look up the contact, then send the mail. The others stop halfway and ask you for the address they were sent to find. The larger E4B variant was rejected on the same grounds: 18.3 tok/s and 1.8 GB more memory for no gain in accuracy.
Installing Gemma 4 E2B from Ollama instead gives you a Q8_0 artifact of about 8.1 GB, which does not fit in 8 GB. The 3.1 GB above is the Q4_0 GGUF this device ships with.

​
Bring your own local model (Ollama)

Measured on a production ClawBox with Ollama (warm runs, num_predict=200, Ollama’s own eval_rate; power and thermals from tegrastats over 176 samples): ~11.4 W average, 19.2 W peak, 61.8 °C peak, no throttling. Most people read at 5–8 tok/s, so the 1–4B class generates faster than you can read. Full methodology: We Benchmarked a Production ClawBox.
For larger models, use your own cloud API key (Claude, GPT, Gemini) — see Choose Your AI Provider. For large models fully local, see ClawBox Workstation (128 GB unified memory, up to ~200B params).

​
Skip the DIY

Get ClawBox — pre-configured, ready in 5 minutes

ClawBox ships with Recommended+ specs, OpenClaw pre-installed, dual-band Wi-Fi 5 (802.11ac) and Bluetooth 5.0. Manual DIY setup is typically 2–4+ hours (plus waiting for hardware).