Configure analysis¶
ARGUS needs one AI provider to investigate anything. Configure it and test it before you enrol clusters, so your first incident can be investigated the moment it appears.
Go to Settings → AI providers.
Configure as many providers as you like. Exactly one is active, and the active one is what every investigation uses. Switching takes effect on the next investigation — there is no restart.
Choose a provider kind¶
| Kind | Use it when |
|---|---|
| API endpoint | You have an OpenAI-compatible endpoint and a key, hosted anywhere including your own infrastructure |
| Agent CLI | You deployed the optional runner beside the hub and want ARGUS to drive an agent CLI |
API endpoint¶
| Field | Value |
|---|---|
| Base URL | Your provider's endpoint |
| Model | The model name as the provider expects it |
| API key | Your key |
Agent CLI runner¶
| Field | Value |
|---|---|
| Runner URL | http://argus-hub-cli-runner:8090, exactly, when the hub chart deployed the runner (http://<release>-cli-runner:8090 for another release name). Not the placeholder shown in the field |
| Agent | claude for an Anthropic API key or a Claude subscription token, codex for an OpenAI key. It must match the credential the runner was given |
| Runner token | The shared token the runner expects. Leave it blank when the hub chart deployed the runner: the hub already has it |
Adding the runner the hub chart deployed: Add a provider → Agent CLI, the runner URL, and the CLI it drives. The runner token stays blank, because the hub already has it.
The runner exists only after you give the hub chart a credential, in Install the hub, Step 3. Do that first; this page is step 4 of that sequence.
Test before you rely on it¶
Select Test connection. A failing test names the reason — an unreachable URL, a rejected key, an unknown model — which is considerably easier to read now than as a failed investigation later.
A passing test for an Agent CLI provider. It proves the runner URL, the runner token and the CLI installation, without calling the model.
A passing test does not prove the credential
For an Agent CLI provider the model credential is first used by a real investigation. Open an incident and select Investigate to confirm it end to end. If that fails with Invalid API key, see Analysis fails immediately.
Bounding cost and time¶
Settings → Investigation bounds every investigation: a maximum number of model round-trips and a wall-clock limit. The defaults are eight iterations and 180 seconds. An investigation that reaches a bound stops and reports what it found so far, marked as partial, rather than continuing indefinitely.
Set a daily spend ceiling under Settings → Model pricing so cost is visible on the overview and bounded in advance. See Model pricing and spend.