Choose providers and models
Agent creation and settings support main, sub, fallback, and heartbeat models. Choices must match actual credentials, the model catalog, and the runtime release.
The current recommendation is to use the same provider and model for main and heartbeat. Check both settings when creating an Agent or changing its models.
Managed providers
| Provider | Credential | Model selection |
|---|---|---|
| Anthropic | API key or setup token | Standard selector; Agent binding accepts API keys only |
| OpenAI | API key | Standard selector |
| OpenAI Codex | ChatGPT sign-in or auth.json | Standard selector; subscription candidates depend on catalog status |
| Google Gemini | API key | Standard selector |
| OpenRouter | API key | Catalog models can appear in the standard selector |
| Local LLM (Ollama) | Server URL and model | Models can be assigned to roles after a connection test; no API key support |
| OpenAI-compatible | Server URL, model, optional API key | Models can be assigned to roles after a connection test |
The Credential dialog has six service choices: Anthropic, OpenAI, OpenAI Codex, Google, OpenRouter, and Local LLM. Local LLM distinguishes Ollama from OpenAI-compatible servers, giving seven runtime provider identifiers. The hosted selector uses the first five hosted providers.
See provider API keys or Codex ChatGPT sign-in. Sign in through the Portal or register an existing auth.json. Follow each service's terms.
Available models
Use the current Portal list rather than a fixed list in documentation. Service and model activation can change. If a model is missing, ask your organization administrator; writing its name in instructions does not enable it.
| Role | Purpose | Selection principle |
|---|---|---|
| main | Primary responses and final decisions | Balance quality, tool compatibility, and cost |
| sub | Research, drafts, and supporting work | Can be lighter than main; do not delegate sensitive judgments |
| fallback | Alternative when main fails | Must have working authentication and access, including another provider's credential if needed |
| heartbeat | Periodic check-in turns | Use the same provider and model as main; this does not enable a Cloud self-service scheduler |
Fallback is a usable alternate route, not simply a cheaper model. If it uses another provider, connect both credentials and verify each in a test Room.
Configure and verify
For a new Agent, select models and credentials during Agent creation. To change an existing Agent:
- Prepare provider credentials.
- Open the target Agent's details from the Agent list and edit its AI resource settings.
- Change providers and models for main, sub, fallback, and heartbeat as needed. We recommend matching heartbeat to the main provider and model.
- Choose a saved key or use the direct key entry field for each provider. Verify the selected credentials are available in the same tenant.
- Check the impact on ongoing work, then choose the action to save and restart. Choose cancel to leave without changes.
- Wait until the Agent is available again, then verify model responses and required tools with a small, non-sensitive test request.
- Track cost, errors, and latency in Analytics. Avoid changing several roles at once when diagnosing problems.
Local LLM prerequisites
Register an Ollama or OpenAI-compatible server and pass the connection test. If models are assigned automatically, review main and heartbeat before saving and match their provider and model according to the current recommendation. Also check access scope, performance, privacy, and fallback.
A successful connection test still requires a real response from the target Agent. Verify server access and model readiness. For other integrations, first confirm a supported Portal setup path; do not guess settings or paste secrets into conversation.