Empty API Key for Ollama blocks run

[vchatela]

Hello,

I used Ollama provider (localhost) but when I set nothing in APIKey configuration when I try to run the agent I get:
instance "sympoz" has no API key configured (authRefs is empty)

I couldn't make it work with Ollama provider (neither with remote Ollama exposed nor local ollama installation).
I would expect a way to validate the AI provider is working because when using Ollama it doesn't show any of the models while I have some available.

$ ollama run qwen3.5:397b-cloud
Connecting to 'qwen3.5:397b' on 'ollama.com' ⚡
>>> 

But on the UI it shows it hasn't find any model.

Regards

[three-foxes-in-a-trenchcoat]

Hey @vchatela 👋

This is a valid bug — Ollama doesn't require an API key, but the controller is enforcing authRefs as mandatory. When you select Ollama during onboarding, the wizard correctly says "No API key needed" but doesn't create a dummy/empty secret, so the controller rejects the run.

Workaround for now: Create an empty secret and reference it:

kubectl create secret generic ollama-dummy-key --from-literal=apiKey=""

Then patch your instance to reference it (or re-run onboard with that secret).

Proper fix needed: The controller should treat Ollama as a special case where authRefs can be empty/omitted. The onboarding wizard should either:

1. Skip secret creation entirely for Ollama and the controller should allow nil authRefs for Ollama provider
2. Or create a placeholder secret automatically

Regarding model discovery — the Ollama provider integration likely needs to call /api/tags on the Ollama endpoint to list available models. If it's not showing models, it might not be reaching the Ollama server (especially if using 127.0.0.1 from inside a pod — you'd need the host network or a proper service endpoint).

I'll look into a fix for this. 🦊

[vchatela]

Thanks @three-foxes-in-a-trenchcoat !

I've simply deployed ollama via their installer on a remote machine and configured http://vm.domain.fr:11434/v1 (with OLLAMA_HOST=0.0.0.0 ) but still not getting models.

Any way to see logs on runnings ? I expect the ollama provider not to work, I've started a run but can't get any log to see what's wrong..

[kristof-ringleff]

I believe there is an additional issue around providing a custom URL. Both with the TUI and the UI it seems any provided URL is not permanently saved. This applies to both Ollama and Custom OpenAI.

This is visible for example in the instances view:

and any runs produces this result:

OpenAI API error (HTTP 401): POST "https://api.openai.com/v1/chat/completions": 401 Unauthorized {

when instead it should have called http://127.0.0.1:4096

[v01t]

401 for custom provider caused by env name mismatch (when instance created from Web UI including creation of secret) between generated env name in secret and expected by agent runner, fixed with #51

[AlexsJones]

Hey all, the core bug here (empty API key blocking Ollama runs) is fixed and verified locally against Ollama with qwen3.5-27b. The createRun path now skips the API key requirement when a baseURL is set and auto-detects the provider from the URL (relevant code), and the controller correctly skips auth secret injection for keyless providers (controller changes). We just tagged v0.5.9 with these changes included, so give it a spin and let us know if you hit anything else or if the model discovery / URL persistence issues from this thread are still a problem on the new tag.