Private AI,
woven into your own hardware.
Open Agent Fabric turns NVIDIA DGX Spark boxes into a complete AI platform. It serves the models, routes requests between them, issues keys, enforces guardrails and records every trace and metric. Your data, prompts and logs never leave your rack.
Proven open source, pinned and run for you.
Upstreams stay unmodified. Open Agent Fabric installs, configures, upgrades and watches each of them.
Models on vLLM
A catalog tuned for GB10. One box, or a cabled pair for larger models. Downloads, placement, health and replacement handled.
→Agent gateway
One OpenAI-compatible endpoint, per-app keys that apps rotate themselves, external providers, and guardrails that reject credentials and mask personal data.
→Semantic routing
model: auto sends simple requests to a fast model and hard ones to a large one. Observe mode first, then active.
→Traces
Every request in Langfuse: who asked, what was sent, which model answered, cost and latency.
→Metrics and logs
Prometheus, Loki and Grafana, with dashboards for every box and model, ready for your apps too.
→Cluster
Join boxes from the console. OAF finds the cables, addresses the 200 GbE links, checks their speed and copies models across.
→Declarative. Idempotent.
One cluster.yaml describes the boxes, the cables, the models and the routing. The console edits that file, and every change is a plan you review before it is applied. Run it again any time; nothing is redone.
spec: topology: pair-plus-single models: - name: fast catalogRef: qwen3-coder-30b-a3b-instruct-fp8 placement: { node: spark-1 } - name: glm-5.3-flash catalogRef: glm-5.3-flash-nvfp4 profile: pair-tp2 # across two cabled boxes placement: { group: pair-1 } routing: mode: observe # watch decisions, then go active
One command. Whole stack.
On a fresh DGX Spark: preflight, models, gateway, observability, and an acceptance test at the end. Every step checks before it acts, so re-running it is safe.
Garrison AI installs on top of it; its installer comes with a licence.
curl -fsSL https://raw.githubusercontent.com/anindyar/openagentfabric/main/install.sh \
| bash -s -- --email you@company.com
# preflight → models → gateway → observability → acceptance testOne open foundation. Use cases on top.
The platform is open. Products for real work are woven on it: Garrison is the first, and we build more with you.
First use case · CommercialGarrison AI
Manage your servers, cloud accounts and Kubernetes with private AI, and keep them secure. Ask Garry in plain language, get answers from read-only checks, approve every change. SIEM alerts, threat hunts and firewalls too.
Your industry, next.
Agents for your domain on the same open foundation: finance, legal, healthcare, manufacturing. Your data stays on your hardware, and you own what we build together.
Talk to us →Build on the fabric.
Run the platform yourself from GitHub, try Garrison on your estate, or build a use case for your industry with us.