OpenClaw on Raspberry Pi / ARM: Everything That Breaks and How to Fix It

The Raspberry Pi is an appealing always-on OpenClaw host: low power draw (~5W), silent, under $100, and you probably already have one. The official docs say it works — with "rough edges." That's underselling it a bit.

ARM64 OpenClaw installs have several known failure modes that the standard setup guides don't cover. This guide documents all of them, with specific fixes for each.

Hardware Compatibility at a Glance

Model	OpenClaw	Ollama Local Models	Notes
Pi 5 (8GB)	✓ Works	✓ Works	Best Pi option. Can run Qwen 2.5 3B comfortably.
Pi 4 (8GB)	✓ Works	⚠ Limited	Ollama works with small models (≤3B). Slow inference.
Pi 4 (4GB)	✓ Works	⚠ Very limited	Only smallest models (1B–2B). API model recommended instead.
Pi 4 (2GB)	⚠ Tight	✗ Not practical	OpenClaw gateway alone uses ~400MB. Very little headroom.
Pi 3	✗ Not supported	✗ No	32-bit OS, ARMv7 architecture. Node v22 requires 64-bit.

Issue 1: The npm Install / Binary Compatibility Problem

The most common ARM failure: you run npm install -g openclaw and get cryptic binary errors like:

Error: /usr/local/lib/node_modules/openclaw/node_modules/better-sqlite3/build/Release/better_sqlite3.node: invalid ELF header

Or:

Error: /usr/local/lib/node_modules/openclaw/node_modules/sharp/build/Release/sharp-linux-arm64.node: cannot open shared object file

Why it happens: Some of OpenClaw's dependencies ship prebuilt native binaries for x86_64. On ARM, npm tries to use those binaries, which are the wrong architecture. Or the ARM64 binary was never built for that dependency version.

Fix — force native compilation:

# Install build dependencies first
sudo apt-get install -y build-essential python3 python3-pip libvips-dev

# Install openclaw and force rebuild of native modules
npm install -g openclaw --build-from-source

This takes 10–20 minutes on a Pi 4 (the compilation is slow on ARM) but produces binaries that actually work.

Issue 2: Node.js Version

Raspberry Pi OS ships with old Node.js versions. OpenClaw requires v20+ (v22 recommended). The system package often gives you v16 or v18.

# Check what you have
node --version

# If it's under v20, install nvm and upgrade
curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.39.7/install.sh | bash
source ~/.bashrc

# Install and use Node 22
nvm install 22
nvm use 22
nvm alias default 22

# Verify
node --version  # should show v22.x.x

Then reinstall openclaw with the correct Node version active.

Issue 3: Skills with External Binary Dependencies

Some ClawHub skills bundle or depend on external binaries that only have x86_64 builds. These silently fail to load on ARM without a clear error message.

Safe skills on ARM64:

• Skills that are pure JavaScript/TypeScript — always work
• Skills using Node.js child_process to call CLI tools you've installed via apt — work if the CLI has an ARM64 build
• WhatsApp via Baileys — pure JS, ARM64 is fine

Potentially broken skills on ARM64:

• Skills that bundle precompiled binaries (check for .node files in the skill directory)
• Skills with native npm dependencies that don't have ARM64 prebuilds

Diagnosis: after installing a skill, run openclaw doctor and check for skill-related warnings. If a skill shows dependency errors, look at its package.json for native deps and check if they have ARM64 support.

Issue 4: Swap Memory for Gateway Stability

On Pi 4 with 4GB RAM, the OpenClaw gateway plus an active conversation can push close to the memory limit — especially with multi-turn sessions building up context. Without swap, you get OOM kills that look like mysterious crashes.

# Check current swap
free -h

# If swap is 0 or very small, increase it
sudo dphys-swapfile swapoff
sudo nano /etc/dphys-swapfile
# Change CONF_SWAPSIZE=100 to CONF_SWAPSIZE=2048

sudo dphys-swapfile setup
sudo dphys-swapfile swapon

Issue 5: Gateway Auto-Start on Boot

OpenClaw doesn't auto-start on reboot by default. On a Pi that's meant to be always-on, this means any power cycle or update requires manual intervention. Here's the systemd service setup:

# Create the service file
sudo nano /etc/systemd/system/openclaw.service

[Unit]
Description=OpenClaw Gateway
After=network.target

[Service]
Type=simple
User=YOUR_USERNAME
WorkingDirectory=/home/YOUR_USERNAME
ExecStart=/home/YOUR_USERNAME/.nvm/versions/node/v22.0.0/bin/openclaw gateway start --foreground
Restart=on-failure
RestartSec=10

[Install]
WantedBy=multi-user.target

# Enable and start
sudo systemctl daemon-reload
sudo systemctl enable openclaw
sudo systemctl start openclaw

# Check status
sudo systemctl status openclaw

Replace YOUR_USERNAME with your actual username, and update the Node path to match your nvm installation (which openclaw will give you the full path).

Issue 6: Ollama on ARM — Model Selection Matters

Ollama runs on ARM64, but the model size ceiling is much lower than on x86 machines:

• Pi 5 (8GB): Qwen 2.5 3B or Llama 3.2 3B work well. Anything 7B+ will be extremely slow.
• Pi 4 (8GB): Same, but slower. Qwen 2.5 1.5B is more practical for interactive use.
• Pi 4 (4GB): Use Anthropic/OpenAI API instead of local models. Not enough RAM for meaningful local inference.

For a Pi-hosted OpenClaw that's meant to be practical, the best architecture is: local Pi for the gateway and workspace, API model for inference. This gives you 24/7 uptime without the API needing to be always running, while still getting Claude or GPT quality responses.