Running OpenClaw on a $10 Raspberry Pi: The Embedded Guide

The mainline OpenClaw requires Node.js, npm, and roughly 390 MB of RAM. That rules out most embedded hardware. But the OpenClaw ecosystem doesn’t stop at modern desktops: three serious rewrite variants target constrained hardware explicitly, and there’s now a dedicated installer that handles seven different OpenClaw flavors on Raspberry Pi hardware. Here’s the practical guide.

Why Run an AI Gateway on Embedded Hardware?

The appeal is straightforward: always-on, low-power, local. A Raspberry Pi Zero 2 W costs around $15, draws about 0.4W at idle, and can run 24/7 for pennies per month in electricity. For a personal AI assistant gateway that monitors messaging channels and responds autonomously, that’s a compelling deployment profile — especially compared to a laptop that needs to stay awake or a cloud instance with monthly billing.

Privacy is the other driver. Running the gateway locally means your message routing, conversation history, and agent configuration never leave your hardware. The model API calls still go to cloud providers (Anthropic, OpenAI, etc.) unless you’re running a local model, but the infrastructure layer stays under your physical control.

The Three Contenders

PicoClaw — sipeed/picoclaw (Recommended for Most)

PicoClaw is the embedded-first OpenClaw variant, built by Sipeed — the company that makes RISC-V development boards. They know embedded hardware, and it shows in the implementation:

RAM: Under 10 MB (99% smaller than mainline OpenClaw)
Binary: Single Go binary, no dependencies
Startup: Under 1 second on a 0.6 GHz single-core processor
Platforms: Raspberry Pi Zero 2 W, RPi 3/4/5, x86_64, ARM64, RISC-V, MIPS

PicoClaw hit 12,000 stars in its first week after launch — faster than any other rewrite variant. The build system includes hardware-specific make targets:

make build-pi-zero        # ARMv7 hard float for RPi Zero 2 W
make build-linux-arm      # Generic ARM 32-bit
make build-linux-arm64    # 64-bit ARM (RPi 3/4/5)

Go’s runtime brings some overhead compared to Rust or Zig, but on a Raspberry Pi Zero 2 W — which has 512 MB of RAM — 10 MB is a rounding error. PicoClaw is the practical choice for anyone who wants solid embedded support without rougher-edged languages.

NullClaw — nullclaw/nullclaw (Minimal Footprint)

NullClaw is written in Zig and achieves numbers that seem implausible until you run it:

Binary size: 678 KB
RAM: ~1 MB
Startup: Under 8ms
Stars: ~6,830

Zig compiles to native code with no garbage collector and no runtime. The result is a binary smaller than many web pages. NullClaw currently supports Linux x86_64, macOS arm64, and macOS x86_64 — ARM Linux (which covers most Raspberry Pi hardware) is still in progress, so check the project’s current release page before committing.

If absolute minimal resource usage is your priority and you’re comfortable waiting for full ARM Linux support, NullClaw is the ceiling for how small this can get. The nullclaw migrate openclaw command imports existing OpenClaw configuration, so switching doesn’t mean starting over.

ZeroClaw — zeroclaw-labs/zeroclaw (Best Security + Performance Balance)

ZeroClaw is the Rust variant from the Harvard/MIT/Sundai community:

Binary size: ~3.4 MB
RAM: Under 5 MB
Startup: Under 10ms (400x faster than mainline)
Platforms: Linux, macOS, Windows; ARM, x86, RISC-V

ZeroClaw is the most security-conscious of the three: device pairing for remote access, execution sandboxing, and allowlisting are built in. It supports 22+ AI providers and includes the zeroclaw migrate openclaw config import tool. For Raspberry Pi deployments where the gateway will have access to sensitive messaging accounts, ZeroClaw’s security defaults are worth the slightly larger binary compared to NullClaw.

The Unified Installer: sanchorelaxo/openclaw-raspberry-installer

If you’d rather not choose yet — or if you want to test multiple variants — the openclaw-raspberry-installer project automates installation for the full range of options on Raspberry Pi 5 (8 GB) with a Hailo 10H AI accelerator:

CLAW_FLAVOR=picoclaw ./install.sh   # Or zeroclaw, nullclaw, openclaw,
                                    # nanobot, moltis, ironclaw

The installer handles multi-flavor support via a single CLAW_FLAVOR environment variable, supports offline install mode, and includes Hailo local inference wiring for on-device model execution. If you’re building on the Raspberry Pi 5 + Hailo 10H AI HAT+, this is your starting point.

Quick Setup: PicoClaw on Raspberry Pi Zero 2 W

Flash Raspberry Pi OS Lite (64-bit) to a microSD card using Raspberry Pi Imager. Enable SSH during setup.
Download the binary from the PicoClaw releases page and transfer it: scp picoclaw-linux-armv7 pi@raspberrypi.local:~/picoclaw
Make it executable: chmod +x ~/picoclaw
Create a config file at ~/.picoclaw/config.json with your API key and platform credentials. If migrating: ./picoclaw migrate openclaw imports from an existing OpenClaw setup.
Run it: ./picoclaw start
Set it to autostart: Create a systemd service unit to restart it on boot.

Total RAM used on a fresh Raspberry Pi Zero 2 W: under 10 MB for PicoClaw plus the OS footprint. A 512 MB board has plenty of headroom.

Mobile: Android Isn’t Just for Phones

The embedded story extends to Android devices. Several ports make OpenClaw (or NanoClaw) work on Android hardware:

AnyClaw (friuns2/openclaw-android-assistant) — native APK, no root, no Termux required
openclaw-termux (mithun50) — Flutter app with built-in terminal, web dashboard, 7 AI providers
phoneclaw (rohanarun) — full 24/7 Android phone automation via side-loaded APK

A retired Android phone on a charger makes for a surprisingly capable always-on AI gateway: WiFi connectivity, a real CPU, and a battery for power resilience. Storage is the limiting factor on older devices, but most use cases fit comfortably under 8 GB.

Which Variant Should You Run?

Raspberry Pi Zero 2 W ($15): PicoClaw. It’s designed for this exact hardware.
Raspberry Pi 3/4/5: PicoClaw or ZeroClaw. Plenty of RAM for either; choose based on security requirements.
Raspberry Pi 5 + Hailo 10H: Use the sanchorelaxo installer. It handles the Hailo inference integration.
RISC-V or exotic embedded hardware: ZeroClaw (explicit RISC-V support) or NullClaw (when ARM Linux support lands).
Android phone: AnyClaw for simplicity, openclaw-termux for the web dashboard.

The embedded OpenClaw ecosystem is moving fast. PicoClaw’s 12,000 stars in one week and NullClaw’s steady growth signal genuine demand for running AI agent gateways on constrained hardware. The tools are good enough to use today — the main limitation is skill/plugin compatibility, since the ClawHub ecosystem was built for mainline OpenClaw’s TypeScript runtime.

If that gap closes — and the rewrite variants are actively working on it — there will be real competition for whether the next wave of AI gateway deployments runs in a cloud container or on a $10 board on your desk.

Recommended Hardware

To follow this guide, you’ll need a Raspberry Pi. These are the recommended options available on Amazon:

Raspberry Pi Zero 2W — The ideal low-power option for PicoClaw. Quad-core 1GHz, 512MB RAM, ~$15. Handles real AI gateway traffic on under 2W power draw.
Raspberry Pi 5 — For heavier workloads or running multiple services alongside PicoClaw. 4-core Cortex-A76, up to 8GB RAM.
MicroSD Card (32GB+) — A fast, reliable microSD card is essential for stable long-term operation.

ClawTrackr participates in the Amazon Associates program. Links above are affiliate links — you pay the same price, and we earn a small commission that helps keep this site running.