PicoClaw on Raspberry Pi Zero W — Deployment Plan¶

Date: 2026-03-19 Hardware: Raspberry Pi Zero W v1.3 (purchased May 2021, unused) Goal: Always-on, near-zero-power AI assistant accessible via Telegram

Overview¶

PicoClaw is an ultra-lightweight AI assistant by Sipeed (<10MB RAM, single Go binary). It connects to remote LLM APIs (Claude, OpenAI, etc.) and interfaces via chat platforms (Telegram, Discord). The RPi Zero W runs it as a thin client — all AI processing happens remotely.

Prerequisites¶

Hardware¶

[x] Raspberry Pi Zero W v1.3
[ ] MicroSD card (32GB recommended, A1/A2 rated — Samsung EVO Select or SanDisk Extreme)
[ ] Micro-USB power supply (5V/1.2A+ — any phone charger or USB port on NAS)
[ ] Micro-USB OTG adapter (for initial setup if needed, optional with WiFi)

Accounts (all free, no credit card)¶

[ ] Telegram bot — create via @BotFather, get bot token
[ ] Groq API key — https://console.groq.com (primary)
[ ] Google Gemini API key — https://aistudio.google.com/apikey (fallback 1)
[ ] OpenRouter API key — https://openrouter.ai/keys (fallback 2)
[ ] Mistral API key — https://console.mistral.ai (fallback 3)

Network¶

Home WiFi SSID and password (2.4GHz only)
Optional: assign static IP in OPNsense DHCP (e.g., 192.168.0.25)
Optional: join Tailscale mesh for remote SSH

Phase 1 — OS Installation (15 min)¶

Option A: DietPi (recommended)¶

Download DietPi ARMv6 image: https://dietpi.com/#downloadinfo
Flash to microSD with Balena Etcher or dd
Before first boot, edit on the SD card:
dietpi.txt: set AUTO_SETUP_NET_WIFI_ENABLED=1
dietpi-wifi.txt: add SSID and password
dietpi.txt: set AUTO_SETUP_LOCALE_NAME=en_US.UTF-8
dietpi.txt: set AUTO_SETUP_TIMEZONE=America/Asuncion
Insert SD, power on, wait ~5 min for first boot
SSH in: ssh root@<ip> (default password: dietpi)
Complete initial setup (change password, minimal install)

Option B: Raspberry Pi OS Lite (alternative)¶

Use Raspberry Pi Imager, select "Raspberry Pi OS Lite (32-bit)"
Configure WiFi, SSH, username in Imager settings
Flash, boot, SSH in

Phase 2 — System Hardening (10 min)¶

# Update system
apt update && apt upgrade -y

# Create picoclaw user
useradd -m -s /bin/bash picoclaw

# Enable tmpfs for logs (reduce SD card wear)
echo "tmpfs /var/log tmpfs defaults,noatime,nosuid,size=30m 0 0" >> /etc/fstab
echo "tmpfs /tmp tmpfs defaults,noatime,nosuid,size=50m 0 0" >> /etc/fstab

# Reduce GPU memory (headless, no display needed)
echo "gpu_mem=16" >> /boot/config.txt

# Disable unnecessary services
systemctl disable bluetooth hciuart triggerhappy

# Reboot
reboot

Phase 3 — PicoClaw Installation (5 min)¶

# Download ARMv6 binary from latest release
PICOCLAW_VERSION="0.2.3"
wget "https://github.com/sipeed/picoclaw/releases/download/v${PICOCLAW_VERSION}/picoclaw_${PICOCLAW_VERSION}_linux_armv6.tar.gz"
tar xzf picoclaw_*.tar.gz
mv picoclaw /usr/local/bin/
chmod +x /usr/local/bin/picoclaw

# Verify
picoclaw --version

# Initialize config
su - picoclaw -c "picoclaw onboard"

Phase 4 — Configuration (10 min)¶

Provider Strategy ($0/month)¶

Anthropic blocked subscription OAuth tokens in third-party agents (Feb 2026). API keys still work but cost $1-25/M tokens. Instead, use free providers with failover:

Priority	Provider	Model	Free Limits	Why
Primary	Groq	Llama 3.3 70B	~14,400 req/day	Fastest (300+ tok/s), GPT-4 quality, tool calling
Fallback 1	Google Gemini	2.5 Flash	250 req/day	Best quality, 1M context
Fallback 2	OpenRouter	Free router	200 req/day	Auto-selects best free model
Fallback 3	Mistral	Small 3.1	1B tokens/month, 2 RPM	Huge quota, slow RPM

Get API keys (free, no credit card)¶

Groq: https://console.groq.com → API Keys → Create
Google Gemini: https://aistudio.google.com/apikey → Create API key
OpenRouter: https://openrouter.ai/keys → Create Key (select free models only)
Mistral: https://console.mistral.ai → API Keys → Create

Edit config¶

Edit ~picoclaw/.picoclaw/config.json:

{
  "agents": {
    "defaults": {
      "model": "groq/llama-3.3-70b-versatile",
      "max_tokens": 4096,
      "temperature": 0.7,
      "max_tool_iterations": 10,
      "restrict_to_workspace": true
    }
  },
  "providers": {
    "groq": {
      "api_key": "<GROQ_API_KEY>"
    },
    "google": {
      "api_key": "<GEMINI_API_KEY>"
    },
    "openrouter": {
      "api_key": "<OPENROUTER_API_KEY>"
    },
    "mistral": {
      "api_key": "<MISTRAL_API_KEY>"
    }
  },
  "channels": {
    "telegram": {
      "bot_token": "<TELEGRAM_BOT_TOKEN>",
      "allowed_users": ["<your_telegram_user_id>"]
    }
  },
  "tools": {
    "web": {
      "provider": "duckduckgo"
    }
  }
}

Model routing (cost optimization)¶

PicoClaw supports rule-based model routing. For heavier tasks, fall back to higher-quality models:

Quick questions → groq/llama-3.3-70b-versatile (fast, free)
Complex reasoning → google/gemini-2.5-flash (1M context, free)
Code generation → mistral/codestral-latest (free, 2 RPM)

Optional: local fallback (internet outage)¶

If you have Ollama on Docker VM:

{
  "providers": {
    "ollama": {
      "base_url": "http://192.168.0.10:11434/v1",
      "api_key": "ollama"
    }
  }
}

Model: ollama/qwen3:8b (~5GB RAM, CPU-only ~5 tok/s, tool calling works)

Phase 5 — Systemd Service (5 min)¶

cat > /etc/systemd/system/picoclaw.service << 'EOF'
[Unit]
Description=PicoClaw AI Assistant
After=network-online.target
Wants=network-online.target

[Service]
Type=simple
User=picoclaw
ExecStart=/usr/local/bin/picoclaw
Restart=always
RestartSec=10
Environment=HOME=/home/picoclaw

[Install]
WantedBy=multi-user.target
EOF

systemctl daemon-reload
systemctl enable picoclaw
systemctl start picoclaw

Phase 6 — Verification (5 min)¶

# Check service status
systemctl status picoclaw

# Check RAM usage
ps aux | grep picoclaw

# Check logs
journalctl -u picoclaw -f

# Test: send a message to your Telegram bot
# It should respond via the configured LLM

Phase 7 — Integration with Homelab (optional)¶

Join Tailscale mesh¶

curl -fsSL https://tailscale.com/install.sh | sh
tailscale up --login-server=https://hs.cronova.dev --accept-dns=false

Add to monitoring¶

Add node-exporter or Telegraf for metrics to VictoriaMetrics
Add Uptime Kuma monitor for PicoClaw health

Guarani name¶

Following the naming convention: Ñe'ẽ (ñe'ẽ = "word/speech" in Guarani) — fitting for an AI assistant that speaks.

Power and Placement¶

Power from NAS USB port (always on, no extra power supply)
Or dedicated 5V/1.2A micro-USB charger
Place near WiFi AP for best signal
No case needed if mounted inside rack/shelf (tiny form factor)

Cost¶

Item	Cost
RPi Zero W	$0 (already owned)
32GB microSD	~$8
Micro-USB cable	~$2 (or reuse)
LLM API (Groq/Gemini/OpenRouter/Mistral)	$0/month (free tiers)
Web search (DuckDuckGo)	$0/month
Total	~$10 one-time, $0/month

Combined free API budget: ~15,000+ requests/day across all providers.

Risks¶

ARMv6 support declining — PicoClaw may drop ARMv6 in future versions
WiFi 2.4GHz only — susceptible to interference
SD card wear — mitigated by tmpfs for logs
Pre-v1.0 software — expect breaking changes between versions
Single point of failure — no redundancy (acceptable for a personal assistant)

Timeline¶

Phase	Time	When
1. OS installation	15 min	At home
2. System hardening	10 min	At home
3. PicoClaw install	5 min	At home
4. Configuration	10 min	At home
5. Systemd service	5 min	At home
6. Verification	5 min	At home
7. Homelab integration	15 min	At home
Total	~65 min

Blocked on: microSD card purchase (~$8).