Connect a cloud model

How providers, keys and the catalog fit together is explained in Models on the phone and in the cloud.

Save a key

1. Open Models from the menu, then Providers and access.
2. Choose the provider — OpenAI, Anthropic, Google Gemini, DeepSeek or OpenRouter.
3. Paste the key and save it.

The key goes into the phone’s Keystore. From then on the app shows only that a key is saved, never the key itself. To change it, paste a new one over it; to remove it, use Remove key.

OpenRouter without copying a key

For OpenRouter you can tap Sign in with OpenRouter instead: the browser opens, you sign in, and OpenRouter issues the key straight to TALOS. If the browser cannot open, or the sign-in fails, paste a key instead.

Use your own Ollama server

Ollama needs no key, only its address:

1. In Providers and access, open Ollama and enter the endpoint.
2. Save it, then refresh its models.

The address must start with http:// or https://, and must not contain a username or password. A plain http:// address is accepted only for a machine on your own network, written as a number — for example http://192.168.1.20:11434. Anything else needs https://.

Each provider also has a timeout, and a custom endpoint can be reset to the official one.

Test a model, then use it

1. Open Model catalog. It lists the models your providers offer, with what each declares it can do — context, inputs, reasoning — kept apart from what was actually observed.
2. Pick a model and run Test completion. Until it passes, the model reads “not tested”.
3. Use model makes it the default; Show in composer puts it in the chat’s quick picker.

The Model Lab and the composer share one selection: change the model in either, and the next answer uses it.

When a model is missing

If a provider does not list a model you know exists, add it under Advanced manual models with its provider and model ID. What it can do — reasoning, image input — is then declared by you, and marked as such.

If it does not answer

• “Add your … API key before continuing” — no key is saved for that provider.
• “… answered with nothing at all” — try another model.
• “… spent its whole token budget thinking” — the model reasoned until it ran out; pick one that reasons less, or a larger one.
• “Ollama requires an endpoint reachable from this device” — the phone cannot reach that address; check that the server listens on your network, not only on its own machine.