This is the smallest useful loop: one GPU workstation, one desk, one prompt.
Fabric is installed on the GPU machine. Agent is installed on the desk. Ollama is running on the GPU host with at least one model pulled. See Installation.
Start the fabric
On the GPU workstation, terminal 1:
houdry serve --listen 0.0.0.0:8080
That is HTTPS. Cluster dashboard: https://127.0.0.1:8080/.
Terminal 2 on that machine (or on a second GPU box):
houdry gpu detect
houdry gpu register
Leave register running. Same command as houdry node join. First join: run houdry node enroll create on the serve machine and use the HDRY_… token on the GPU host.
Connect Houdry Agent
Open Houdry Agent. Choose Houdry server URL.
| Agent is on… | URL |
|---|---|
The same machine as houdry serve | https://127.0.0.1:8080/v1 |
| Another machine on the LAN | https://HOST:8080/v1 |
/v1 is required. Model: auto. Optional key: houdry if you passed --token to serve. On one control plane on Wi-Fi, the URL often fills in. Confirm with Connect.
Put your files in Knowledge
In Agent, open Knowledge base. Add the SOPs, inspection reports, and manuals this site is allowed to read.
- Files stay on this computer. They are not uploaded to a public model host.
- On each SOP, set standing rules (cite the SOP, do not invent procedure, stay on this network).
Do not paste whole manuals into chat if they already live in Knowledge.
Run one real task
Start a New session. Keep the first prompt narrow:
Read 01_Inspection_Report_E204.pdf and the matching SOP.
Draft an approval note. Quote every finding from the source. Do not invent values.
The fabric routes the work (model=auto) to a READY GPU node. The answer comes back in Agent. Nothing is mailed out.
If chat does not answer
houdry discoveron the GPU machine: is the control plane visible?houdry gpu detect: is a GPU listed?- Is
houdry gpu registerstill running (not exited)? - Is Ollama up, with a model pulled?
- Is Agent pointed at
https://HOST:8080/v1, not a public LLM URL, and not HTTP?
More in Troubleshooting.