Skip to content

Ollama

Ollama is optional. BarkVisor talks to it on each Device. Completions go through Home :7777/v1, not Device :11434.

The Ollama nav item stays visible for admin and inference even when no Device can reach Ollama.

The page shows a multi-step panel and Recheck. Commands match the in-app copy. AgentBox and Mac mini are not install targets.

macOS

Terminal window
brew install ollama
brew services start ollama

Linux — install the distro package, or see ollama.com/download.

When Ollama comes up, the page switches to the model list without a full reload.

Live Device stats (CPU, memory, GPU busy percent) are on Device detail, not this page. Occupancy / passthrough is a different number.

Filter catalog matches names already pulled on the Home. That is not a library search.

Search the Ollama library looks up names through Home (GET /api/home/ollama/library/search?q=). The daemon fetches https://ollama.com/api/tags (allowlisted) and filters locally. Empty query, upstream down, and no matches are explicit states. Each result has Download, which is the existing pull (POST /home/ollama/pull).

Pull by name stays as a fallback when you already know the slug.

Pull still needs a Device for where the weights land (sidebar Device if one is selected, otherwise any reachable Device).

A pulled model can only run on a Device that already has it.

  • Sidebar scoped to one Device that has the model: Start runs there. No picker.
  • Sidebar All, one reachable location: Start runs there. No picker.
  • Several reachable Devices have it: picker lists only those Devices. Always a real Device, never “Any reachable Device.”
  • No reachable location: Start is disabled (not on this Device, or only on unreachable Devices).

Stop uses the Device that is running the model and asks for confirmation first.

Home serves OpenAI-compatible completions at:

http://<home>:7777/v1/chat/completions

That is not Device :11434. Send Authorization: Bearer with an inference key from Settings.