Houdry splits compute from the people using it.
On the desk
Houdry Agent
Chat, Knowledge, SOPs, files, and local tools.
Talks to Fabric at http://HOST:8080/v1
On the GPU
Houdry Fabric
houdry serve on port 8080 is the control plane.
houdry gpu register runs Ollama and returns the answer.
People never talk to the GPU process. They talk to Agent. Agent talks to Fabric. Fabric picks a model on a READY GPU.
Two Fabric processes, plus the desktop app
| What you run | Product | What it does |
|---|---|---|
houdry serve on port 8080 | Fabric | HTTP API, dashboard, LAN announce, router |
houdry gpu register | Fabric | Stay READY, claim jobs, run models |
| Houdry Agent | Agent | Chat, sessions, knowledge, local tools |
houdry node join is the same worker as gpu register. houdry gpu join is a one-shot inventory snapshot and does not run jobs.
One machine can run serve and register together. Extra GPU boxes only run register.
What a request does
Step 1
Prompt
Operator types in Houdry Agent.
Step 2
LAN call
POST /v1/chat/completions with model auto.
Step 3
Route
Fabric scores the task against READY GPUs.
Step 4
Infer
That workstation runs Ollama and streams tokens.
Step 5
Reply
The answer shows up in Agent. Files stay on-prem.
Analyze is a heuristic (keywords, length, images, tools), not a second LLM call. Scoring prefers a loaded, right-sized model on a READY node. Vision attachments go to a vision-capable model when one is available.
If no GPU is registered yet
With no READY GPU node, chat can still talk to Ollama on the serve host. When a GPU registers, the same Agent URL starts dispatching cluster jobs. You do not change the Agent URL.
Agent settings
| Setting | Typical value |
|---|---|
| Provider | Houdry server URL |
| Base URL | http://HOST:8080/v1 |
| Model | auto (fabric router) |
| API key | houdry if serve was started with --token |
Same machine as Fabric: http://127.0.0.1:8080/v1. Browser dashboard is http://HOST:8080/ (no /v1). Agent needs /v1.
Data directory:
- Linux / macOS:
~/.houdry-agent - Windows:
%LOCALAPPDATA%\houdry-agent
Fabric state lives under ~/.houdry (HOODRY_HOME overrides). Generated files can be served from /files/.
Network
- Fabric announces on the LAN (
_houdry._tcpand UDP41808). - Agent and extra GPU nodes browse for that announce.
- Guest Wi-Fi that isolates clients blocks discovery. Paste a URL instead.
- There is no cloud LLM hop. Pull the WAN cable; chat against local models still works if Ollama and Fabric are up.
- Do not bind
houdry serveto the public internet.