Raspberry Pi 5 deployment¶

The Pi 5 isn’t just a deploy target — it’s the numerical audit rig for llamaclaw. Every function in esml is tested on both x86_64 and aarch64; subtle BLAS differences catch ill-conditioned code that x86 Accelerate silently accepts.

Hardware recommendation¶

Raspberry Pi 5 (8 GB minimum, 16 GB preferred for Perseus)
Active cooler (official kit works)
1 TB NVMe via PCIe HAT — puts /home and Ollama models off the SD
Ethernet (Wi-Fi works but NVMe downloads are large)

1. Base OS¶

Flash Raspberry Pi OS 64-bit (Bookworm or later). First boot: set hostname to zeus.local, create user perseus, enable SSH.

2. Move /home to NVMe¶

See llamaclaw/deploy/yodavision/zeus/APPLY_NOTES.md for the full migration. SD card stays boot + root; all user data + Ollama models live on NVMe.

3. Bootstrap¶

ssh perseus@zeus.local
curl -fsSL https://install.llamaclaw.org | bash -s -- --with-pi

The --with-pi flag pulls Perseus (llamaclaw/perseus:e2b), wires up systemd units from llamaclaw/deploy, and runs kernel-tuning udev rules (NVMe BFQ scheduler, sysctl tweaks).

4. Verify¶

ssh perseus@zeus.local
esml doctor
# ─── ESML Doctor ───
# ✓ Python 3.12.3
# ✓ R 4.4.1
# ✓ Ollama 0.1.40
# ✓ Perseus perseus:e2b (9.6 GB)
# ✓ 41 datasets in SQLite
# ✓ NVMe /home mounted (1 TB, 620 GB free)
# ✓ Luci systemd unit enabled

5. systemd units¶

Three services come from llamaclaw/deploy/raspberry-pi/:

Unit	What
`esml-ollama.service`	Keeps Ollama daemon running (restart-on-failure)
`perseus-relay.service`	Perseus HTTP API at port 8421 for network access
`luci.service`	Optional: Luci AI agent (see below)

Enable all three:

sudo systemctl enable --now esml-ollama.service perseus-relay.service

6. Talk to Luci (optional)¶

Luci is the Pi-side AI agent from llamaclaw/luci. It’s sandboxed: no agent-to-agent calls, no public endpoints, exec approvals required.

# Interactive
ssh -t perseus@zeus.local "zeroclaw agent"

# One-shot
ssh perseus@zeus.local "zeroclaw agent -m 'what is your status?'"

7. Run llamaclaw/tide remotely¶

ssh -t perseus@zeus.local "docker run --rm -it --network host ghcr.io/llamaclaw/tide:latest"

Or build kronos natively (ARM-first, ~8 MB binary):

ssh perseus@zeus.local
git clone git@github.com:llamaclaw/kronos.git && cd kronos
cargo build --release
./target/release/kronos

Security posture¶

SSH key-only (no password auth). Deploy key on ruhelavansh-oss/esml is separate from the user-level key on llamaclaw (different auth paths for the monorepo mirror vs the new ecosystem).
ufw firewall active; only SSH + Perseus-relay port exposed.
Luci runs as dedicated Linux user luci (not root, not perseus).
All systemd units use Restart=on-failure with exponential backoff.

Pi isn’t just for deployment¶

Every llamaclaw CI run should eventually be reproduced on a Pi, because aarch64 BLAS raises LinAlgError on matrices that x86 silently tolerates. See yodavision/mistakes/arm_vs_x86_numerical_divergence.md for the pattern we discovered on 2026-04-17: two pytest failures on Pi that never showed on Mac, both caught by ARM’s stricter numerics.