A complete open-source AI agent framework designed to run entirely on local hardware — no cloud required, no compromises.
The openclaw hardware software stack is built on a modular architecture that separates concerns cleanly while allowing deep integration. At its heart is the OpenClaw agent framework — a Node.js runtime that orchestrates language models, tools, and plugins into a cohesive AI assistant experience.
Every component in the stack is open source, auditable, and runs locally. When you build a DIY ClawBox build, you're not just assembling hardware — you're deploying a complete software platform that you can inspect, modify, and extend however you like.
The most critical software choice for any desktop AI openclaw build is the LLM inference engine. The OpenClaw framework supports multiple backends, and you can switch between them depending on your hardware and model preferences:
For most builders, Ollama is the recommended starting point. It provides a simple CLI and REST API for downloading, managing, and running language models. With a single command — ollama run llama3.1 — you have a fully functional local LLM. Ollama handles quantization automatically, manages GPU offloading on Jetson and x86 systems, and maintains a curated model library. The OpenClaw framework connects to Ollama's API seamlessly, making it the path of least resistance for your openclaw AI assistant.
For maximum performance on ARM devices like the Raspberry Pi 5, llama.cpp is the gold standard. Its aggressive CPU optimizations — including NEON SIMD instructions on ARM — squeeze every drop of performance from modest hardware. The GGUF model format supported by llama.cpp is the most widely available quantized format, giving you access to thousands of pre-quantized models on Hugging Face. Advanced features like speculative decoding and KV cache quantization help even low-end hardware punch above its weight.
Jetson builders can leverage NVIDIA's TensorRT-LLM for maximum GPU acceleration. It compiles models into highly optimized inference engines that exploit every GPU core, achieving 2-4x throughput improvements over generic runtimes. This is the professional-grade option for builders who need fast, production-quality inference.
Voice interaction transforms a silent box into a conversational AI companion. The OpenClaw voice pipeline is fully local and privacy-preserving:
The voice pipeline is optional — your ClawBox works perfectly as a text-only assistant, and many users interact primarily through a web interface, Telegram bot, or terminal.
The true power of the openclaw private AI platform comes from its extensibility. The plugin system allows your assistant to interact with the real world:
The OpenClaw software runs on standard Linux distributions. For Raspberry Pi builds, we recommend Raspberry Pi OS (64-bit, Bookworm or newer) or Ubuntu Server 24.04 LTS. Both provide excellent ARM64 support, stable kernels, and wide peripheral compatibility. For x86 systems, Ubuntu Server or Debian are excellent choices.
Installation is simple: a single bootstrap script handles everything from dependency installation to model downloading. The script is fully auditable — every package it installs and every configuration it touches is documented. For advanced users, a Docker-based deployment is also available, providing containerized isolation and easy updates.
Once installed, your openclaw edge compute system runs as a systemd service, starting automatically on boot and persisting through power cycles. The web-based configuration interface lets you tweak every setting without editing config files by hand — though raw config files are always available for those who prefer them.
Full setup instructions in the Software Setup Guide.