⚠️ This is a mirror — the canonical repo lives on GitLab

👉 https://gitlab.com/ncz-os/meta-base

Source, releases, issues, merge requests, and CI all live on GitLab. This GitHub copy is a read-only mirror and may lag. Please file issues and get releases there.

---

📍 Moved to GitLab

The canonical, authoritative home of this project is GitLab — always:

👉 https://gitlab.com/ncz-os/meta-base

This GitHub repository is a frozen, read-only mirror. All development, issues, and releases happen on GitLab. Please open issues and merge requests there. The full history of this stub is preserved on GitLab.

---

meta-nclawzero-base

Yocto layer for building minimal Linux images for the Raspberry Pi family, plus optional AI-agent runtime recipes layered on top.

All recipes are ARM-universal (COMPATIBLE_HOST = "(aarch64|arm).*-linux"). The maintainer has booted Pi 4 (2 GB + 8 GB); other family members (Pi 5, Pi 3 64-bit, Pi Zero 2 W) are recipe-compatible but untested. See INSTALL.md for the full MACHINE value matrix with tested/untested flags.

nclawzero currently supports:

• x86_64 + macOS via Docker/Podman containers (ghcr.io/perlowja/nclawzero-demo, ghcr.io/perlowja/nclawzero-agent)
• ARM Raspberry Pi family (Pi 4, Pi 5, Pi Zero 2 W, Pi 3 64-bit) via Yocto-built flashable images from this layer or pre-built SD images from pi-gen-nclawzero

Jetson family support is deferred pending hardware validation.

No pre-built images are distributed from this repository. Everything is built locally from upstream source and this layer's recipes.

What's in here

Base

• nclawzero-base-image — headless console image: systemd, SSH, networking, shell/dev tools, Python. This is the "just Linux + goodies" image. No compositor, no agent, nothing else.
• nclawzero-desktop-image — base + the Weston Wayland compositor with its built-in VNC backend. Connect from your workstation via vncviewer <device-ip>:5901.

Optional AI agent stack (opt-in)

• nclawzero-agent-image — base + ZeroClaw (Rust AI agent runtime)
  • NemoClaw (Node.js sandbox framework, provisioned from upstream github.com/NVIDIA/NemoClaw on first boot).
• packagegroup-nclawzero-agent — same dependencies, for inclusion in your own image recipes.

Recipes

Recipe	Role
recipes-core/images/nclawzero-base-image.bb	Headless minimal image
recipes-core/images/nclawzero-desktop-image.bb	Base + Weston/VNC
recipes-core/images/nclawzero-agent-image.bb	Base + agent stack
recipes-core/packagegroups/packagegroup-nclawzero-base.bb	Shell/dev tools
recipes-core/packagegroups/packagegroup-nclawzero-agent.bb	AI agent runtime
recipes-core/packagegroups/packagegroup-nclawzero-desktop.bb	Weston, mesa, fonts
recipes-connectivity/nclawzero-network/	Wired networking defaults
recipes-devtools/nodejs/	Pre-built Node.js aarch64 binary
recipes-zeroclaw/	ZeroClaw agent (pulls public GitHub release)
recipes-nemoclaw/	NemoClaw (builds from upstream + local patches)

Licensing

This layer is Apache-2.0 licensed. Upstream projects retain their own licenses:

• ZeroClaw (zeroclaw-labs/zeroclaw) — MIT
• NemoClaw (NVIDIA/NemoClaw) — Apache-2.0

The recipes-nemoclaw/ tree contains three local patches authored by Jason Perlow. Those patches are themselves Apache-2.0 derivative work against upstream NemoClaw:

• 0001-fix-snapshot-symlink-protection.patch — security: guard cpSync and directory walker against symlink traversal.
• 0002-fix-config-file-permissions.patch — security: enforce mode 0600 on credential config files.
• 0003-feat-agent-defs-zeroclaw.patch — integration: register ZeroClaw as an agent runtime alongside Hermes.

No binaries from any vendor are redistributed through this layer. All fetches happen at build time from public upstream sources.

Layer dependencies

• openembedded-core (poky)
• meta-openembedded/meta-oe, meta-openembedded/meta-python, meta-openembedded/meta-networking
• meta-raspberrypi (for Pi targets)

Getting started

• QUICKSTART.md — one-page TL;DR: hardware requirements (host + target), a six-step build recipe, pointers into the upstream Yocto Project documentation. Start here if you've never used Yocto before.
• INSTALL.md — full build-host setup, bblayers.conf / local.conf templates.
• FLASH.md — writing images to SD cards on Raspberry Pi (bmaptool, dd, macOS-specific notes).
• docker/README.md — Docker deliverables: nclawzero-demo (one-command interactive bench with zterm TUI + live agent) and nclawzero-agent (headless runtime for compose/k8s). Published to ghcr.io/perlowja. If you just want to feel the stack before committing to hardware, this is the fastest path.

New to Yocto? The Yocto Project Quick Build is a 20-minute walkthrough that builds core-image-minimal for qemu — everything meta-nclawzero-base does is plain Yocto on top of that foundation.

Try the stack in 30 seconds (Docker)

Before touching Yocto or flashing hardware, you can kick the tires on any x86_64 / arm64 machine with Docker installed:

docker run -it --rm ghcr.io/perlowja/nclawzero-demo

That's it — zterm TUI opens, already connected to a live ZeroClaw agent inside the container. See docker/README.md for the nclawzero-agent headless variant.

Layer compatibility

• Yocto release: scarthgap (5.0 LTS).
• Tested build hosts: Debian 12, Ubuntu 24.04.

Contributing

Issues and PRs welcome at this repository. Upstream fixes to NemoClaw or ZeroClaw should be sent to their respective repositories directly:

• github.com/NVIDIA/NemoClaw
• github.com/zeroclaw-labs/zeroclaw

Author

Jason Perlow — <jperlow@gmail.com>