Running a Nodofi AI Worker

Run an AI model on your own machine with Ollama and make it available over the relay — to Studio's Agentic AI and Assistant, robot missions, Flows and published services. No public address, no open port.

Overview

The AI Worker is a small Node.js program that runs next to Ollama. It connects out to a Nodofi relay under a name (for example JozieAILLM2) and answers prompts sent to that name. Ollama only ever talks to the worker on localhost; it is never exposed to the internet or a browser.

Studio / a Flow / a robot ──▶ relay ──▶ AI Worker "JozieAILLM2" ──HTTP──▶ Ollama (gemma4)
                          ◀── answer ◀──

Download the AI Worker

What uses it

Who	What it asks	Worker name
Studio's Agentic AI	Its planner: which Flow or robot skill to run next (Agentic AI)	`JozieAILLM2` — the Studio planner agent
Studio's AI Assistant	Code help in the editor (Assistant settings → Ollama)	`NodofiOllamaCoder` by default
Robot missions, cloud planner	Each step of a mission (Missions on the robot)	the name set in Mission on a robot
Published services	`JozieAI.ask(name, prompt, callback)`	any
Flows	The Call relay worker block with `{ "function": "prompt", "prompt": "…" }`	any

To serve both the planner and the Assistant, run two copies with different WORKER_NAMEs; they can share one Ollama.

Install & run

Install Ollama and pull a model: ollama pull gemma4 (ollama list shows what you have).
Extract the download (or use the ollama-coder folder of the Studio repository) and run npm install (Node.js 18+).
Copy .env.example to .env; set RELAY_URL and WORKER_NAME.
Run npm start. It prints Connected to <relay> — joining as "<name>".
In Studio's Agentic AI pane set the same Relay URL, Studio planner agent and Planner model, and try a goal.

Run it on whichever machine has the GPU (a desktop, a server, a Jetson); it only needs outbound access to the relay.

Start at boot

# /etc/systemd/system/nodofi-ai-worker.service
[Unit]
Description=Nodofi AI Worker
After=network-online.target
Wants=network-online.target

[Service]
WorkingDirectory=/opt/nodofi-ai-worker
ExecStart=/usr/bin/node src/worker.js
Restart=always

[Install]
WantedBy=multi-user.target

sudo systemctl enable --now nodofi-ai-worker. On Windows, use a service wrapper such as NSSM, or Task Scheduler.

Configuration (.env)

Setting	Default	Description
`RELAY_URL`	required	The relay to join — the same one Studio, your Flows and robots use.
`WORKER_NAME`	`NodofiOllamaCoder`	The name callers reach it by; unique on the relay.
`OLLAMA_HOST`	`http://127.0.0.1:11434`	Where Ollama runs. Keep it on localhost unless you protected a remote Ollama yourself.
`OLLAMA_MODEL`	`gemma4`	The model when a request doesn't name one (Studio's planner and Assistant always name theirs).

Protocol

A request is a relay message to the worker's name; the reply goes back to the sender.

`function`	Fields	Reply
`prompt`, `chat`	`prompt` (text, or `[{ role, content }]`), `model`, `format` (`"json"`: Ollama's JSON mode), `think` (`false`: no reasoning pass)	`{ result: true, content, thinking? }`
`generate`	`prompt`, `model`, `think`	`{ result: true, content }`
`embedding`	`prompt`, `model` (e.g. `embeddinggemma`)	`{ result: true, embedding }`

const reply = await NodoRequestAsync('JozieAILLM2', {
  function: 'prompt', functionType: 'async',
  prompt: 'Summarize: …', model: 'gemma4', think: false
}, 'https://relay.nodofi.com');
console.log(reply.content);

Errors reply { result: false, message }. With a thinking model such as gemma4, think: false is several times faster; Studio's planner also sends format: "json" so its replies are always valid JSON.

Monitoring

On a Studio Design map, add an AI Worker component with the worker's name as its Address / ID and press Monitor: it is green while the worker is connected to the relay and red when it isn't — see Monitoring your network.

Troubleshooting

The Agentic AI says the planner "did not answer"

The worker isn't running, or it joined a different relay or name than the one set in the Agentic AI pane. Its log shows joining as "…".

"Could not reach Ollama"

Ollama isn't running at OLLAMA_HOST, or the model isn't pulled (ollama list).

"Relay rejected this worker name"

That name belongs to a registered node on the relay; choose another WORKER_NAME.

Slow answers

Use a smaller model or a GPU machine, and send think: false with thinking models.