NewFire

NewFire is a two-machine AI homelab that hosts the NewFire AI platform. It pairs a Minisforum X1 Pro 370 (control plane) with an NVIDIA DGX Spark (GPU compute) to run local LLMs, agent orchestrators, and tenant-aware API metering for production workloads.

Target launch: May 1, 2026.

Architecture at a Glance

Role	Machine	Specs	Primary Services
Control Plane	Minisforum X1 Pro 370	Ryzen AI 9 HX 370, 96 GB DDR5, 2 TB NVMe	OpenClaw, APISIX, OpenHands, OpenCode, Ollama (CPU)
Compute Engine	NVIDIA DGX Spark	GB10 Grace Blackwell, 128 GB unified, 4 TB NVMe	NemoClaw, Ollama (GPU), vLLM, model workers

The two nodes connect over 1 Gbps LAN (192.168.1.157 and 192.168.1.158) and are remoted via Tailscale plus a self-hosted zrok2 + OpenZiti overlay. OpenRouter is wired in as a fallback cloud LLM gateway.

See AI_Homelab_Architecture.png / .svg / .pdf for the full diagram.

Documentation

The numbered docs in this repo walk through the build in order:

1. 00_OVERVIEW.md Architecture overview, hardware roles, software stack, network layout
2. 01_DGX_SPARK_RECOVERY.md Recovery procedure for the DGX Spark node
3. 02_MINISFORUM_UPGRADE.md Minisforum upgrade and hardening steps
4. 03_INTEGRATION_PATTERNS.md Cross-node integration patterns (ACP, webhooks, message bus)
5. 04_DGX_SPARK_SETUP.md Fresh setup for the DGX Spark
6. 05_OPENROUTER_INTEGRATION.md OpenRouter proxy and key management
7. 06_APISIX_METERING.md APISIX gateway with per-tenant metering
8. 07_RESOURCE_ALLOCATION.md CPU, RAM, and VRAM budgets per service
9. 08_CHECKLIST.md Pre-launch checklist

Supporting material:

• NewFire_Gap_Map.svg, NewFire_Gap_Graph.svg Eight-layer framework gap map (data and no-code builder are the largest current gaps)
• blueprint/ Canonical AI Operating System blueprint (8 layers, 5 agents, 7 maturity levels)
• progress/ Per-initiative progress logs that resume across sessions
• scripts/ Operational scripts for both nodes
• newfire_backend_docker/ Backend service that coordinates Paperclip, OpenClaw, APISIX, and NemoClaw with smart routing
• newfire-db-backup-20260417.sql Database snapshot from 2026-04-17

Software Stack

Component	Purpose	Host	Port
OpenClaw	Multi-agent orchestrator and gateway	Minisforum	18789
OpenHands	Browser-based AI dev agent	Minisforum	3000
OpenCode	AI coding agent	Minisforum	3002 / 3030
Ollama	Local LLM serving	Both	11434
vLLM	High-throughput GPU inference	DGX Spark	varies
NemoClaw	Tenant isolation and model management	DGX Spark	varies
APISIX	API gateway with metering	Minisforum	9080 / 9443
SIE	Embeddings and reranking (bge-m3, bge-reranker-v2-m3)	Minisforum	8089
OpenRouter	Cloud LLM fallback gateway	Cloud	n/a

Current Status

The platform is live at newfire.app with four real tenants onboarded. Active workstreams include the NewFire Sandbox Service (NSS) for isolated agent runtimes, OpenClaw v1 as the coordinator surface, and vLLM tuning on the GB10 for the qwen3-coder-30b NVFP4 build.

Security

Every service is fronted by Tailscale or zrok2 + OpenZiti. No service is exposed on the public internet without an authenticated tunnel. Backups are encrypted, tenant data is scoped, and the immigration-law tenant runs against local-only models.

CI

GitHub Actions runs lint (ruff) and pytest on every push and PR against main — see .github/workflows/ci.yml. It runs on GitHub's hosted runners, so it only checks code, not against anything on the homelab network.

There's also a local review helper, scripts/ci_review.py — run it before opening a PR and it diffs your branch against origin/main and asks the DGX's own qwen3-coder-30b model (over Tailscale) for a second look. Nothing external, no API key, no keeping the diff anywhere but your own machine and the DGX.

License

Not yet specified. All material in this repository is proprietary to the NewFire project until a license is added.