Ollama¶

The Ollama connector watches a local Ollama server: is it up, which version, which model is loaded, and how fast the last measured request answered.

It needs no API key. The account row says local API - no key required; that is expected, not a lost credential.

Set it up¶

Open the admin page and choose Add source -> Ollama.
Give the account a name.
Set base_url to the Ollama server, for example http://192.168.1.50:11434.
Press Save source.

What it reads¶

Endpoint	Used for
`GET /api/version`	Reachability and version
`GET /api/tags`	The models available locally
`GET /api/ps`	The models currently loaded in memory

All of them are plain GET requests. No prompt is ever sent.

The honest limitation¶

Ollama exposes no request history. There is no prompt log, no token history and no duration history. The API tells you which models exist, which is loaded right now, the version, and the timing fields of a request you just made.

So "tokens used today against Ollama" cannot be reconstructed from Ollama alone. Polling its health tells you it is up, not what it did.

The dashboard therefore reports reachability, the version, the loaded model and the number of available models. The speed is a different story.

The measured speed¶

If a collector has written a samples file, the dashboard reads the last sample and computes tokens per second from it.

From the Ollama documentation:

To calculate how fast the response is generated in tokens per second (token/s), divide eval_count / eval_duration * 10^9.

All Ollama durations are in nanoseconds.

Set samples_file to the JSON file such a collector writes. When no readable sample exists, the card reads last measured: unavailable. It does not guess, and it does not show zero.

The dashboard never prompts the model

The speed is a number someone else measured. The daemon sends no prompt and makes no inference request, so it never spends your compute.

Failure behaviour¶

An unreachable server makes the account offline and keeps the last known readings, visibly stale. A reachable server with an unexpected response degrades that one metric to unavailable and leaves the rest alone.