Ollama (local AI)

Point the assistant at a locally hosted language model through Ollama.


Category	AI
Authentication	None
Reaches	Your Ollama host
Needs an agent	No, if the host is reachable from the deployment
Demo mode	Yes, on by default

Capabilities

ai.chat

What it changes, and what it does not

The assistant works without this connector. A built-in command engine handles the recognised commands offline, with no external service and no usage cost. That always works.

Adding a model makes it more flexible. The model maps unusual phrasings onto the same commands, so “who’s got laptop loans out right now” works as well as the canonical wording.

The model never gains new powers. It selects among the same commands, subject to the same permissions, with the same confirmation prompts before anything changes data. It is a better front door, not a wider one.

Nothing leaves your infrastructure

The model runs on a machine you control. Nothing is sent to OpenAI, Anthropic, Google or anyone else.

That is the reason this connector is Ollama rather than a hosted API. A service desk assistant sees student names, device serials and repair notes, and for a lot of schools sending that to a third-party model is not a decision the ICT team can make on its own.

If that distinction matters to your privacy assessment, and it usually does, the compliance page states it.

Setting up Ollama

Install Ollama on a machine with enough memory. A modern desktop with 16 GB works for the smaller models; a GPU makes it pleasant rather than merely usable.
Pull a model:
Terminal window
```
ollama pull llama3.1
```
Confirm it answers:
Terminal window
```
curl http://localhost:11434/api/tags
```
If Plugboard is on a different machine, make Ollama listen on the network rather than only on localhost, and restrict access to the Plugboard host at the firewall. Ollama has no authentication of its own.

Configuring it in Plugboard

Admin, Connectors, Ollama (local AI), Configure.

Field	Default	Value
`baseUrl`	`http://localhost:11434`	The Ollama host
`model`	`llama3.1`	A model you have pulled
`demoMode`	`true`	Skip the model and use the built-in command engine only

There are no credentials.

demoMode is on by default, so installing the connector does not silently start sending anything anywhere. Turn it off when Ollama is actually reachable.

Which model

Bigger is better at understanding awkward phrasing, and slower. The job here is narrow: map a sentence onto one of a known set of commands. A small instruct-tuned model does it well.

Start with llama3.1 at 8B. If responses feel slow, try a smaller model before buying hardware. If the assistant frequently misunderstands, try a larger one.

Private addresses

Plugboard blocks outbound requests to private address ranges by default, so a tenant-supplied URL cannot be used to reach internal hosts.

An Ollama host on your LAN is a private address. A self-hosted deployment needs:

ALLOW_PRIVATE_EGRESS=1

See what you need.

Troubleshooting

Symptom	Cause
The assistant still only handles exact phrases	`demoMode` is on, or the connector is not enabled
Connection refused	Ollama is not running, or is bound to localhost only on another machine
Blocked by the egress guard	`ALLOW_PRIVATE_EGRESS` is not set on a self-hosted install
`model not found`	The model has not been pulled on that host. Run `ollama pull`
Very slow responses	The model is too large for the hardware. Try a smaller one
Answers are confidently wrong	The model is choosing the wrong command. The confirmation prompt before any change is why this is annoying rather than dangerous

The assistant for what it can actually do, and the MCP server for the same catalogue exposed to an external AI client.