Hyperspace Pods let multiple machines pool their GPUs into one private
P2P mesh AI cluster. Models are split across all connected GPUs —
e.g. two machines with 16GB VRAM each can run Qwen 3.5 32B together.
Changes:
- Add modules/nixos/services/hyperspace.nix — NixOS module that:
* Fetches the Hyperspace CLI binary (v5.45.30) via fetchurl
* Sets up systemd service for the agent
* Opens firewall ports (libp2p 4001, chain 30301, API 8080)
* Configures GPU passthrough for AMD MI50 (ROCm)
- Register module in flake.nix for lazyworkhorse
- Enable hyperspace service on lazyworkhorse (ai-worker user, port 8080)
Usage after deployment:
hyperspace pod create "tdnde-lab" # create pod
hyperspace pod invite # share invite with cyt-pi
curl http://localhost:8080/v1/chat/completions # OpenAI API
See skill: nixos-hyperspace-pods
- Add headless openclaw node systemd service for host execution
- Migrate from nix to lix package manager
- Permit openclaw-2026.3.12 (insecure package warning)
- Use ai-worker user for node service
- Started OpenCode service and verified it's running
- Tested Context7 web search functionality
- Tested DuckDuckGo web search functionality
- Documented web search integration in open_code_server.nix
- Updated ROADMAP and STATE with completion status
- Phase 4 complete, ready for Phase 5: TAK Server Integration
