Providers and model roles
Connect intelligence once, choose models per activity, and understand optional services.
Providers are configured once per gateway under Settings → Providers. All agents use that installation's available connections. Adding a provider to the desktop's local gateway does not also add it to a remote server gateway.
Available connections
| Connection | Credential or setup |
|---|---|
| Anthropic | API key |
| Claude Code | Token obtained through claude setup-token |
| OpenAI | API key |
| Codex | The device sign-in flow offered in Jian |
| OpenRouter | API key |
| Gemini | Google Gemini credential |
| xAI | API key |
| Groq | API key; also useful for supported audio tasks |
| OpenAI-compatible | Reachable endpoint and configured credential |
The app's catalog determines the models currently available for each connection. Provider plans, limits, model capabilities, and permitted use remain the provider's responsibility. Jian does not include a model subscription.
The gateway keeps one Anthropic credential at a time: connecting Claude Code replaces the Anthropic API-key connection, and connecting an API key replaces the Claude Code token.
Model defaults by activity
Settings → Model defaults controls separate activities:
| Activity | Purpose |
|---|---|
| Conversation | Chats and default API/channel responses |
| Compaction | Summarize older context when conversations grow |
| Image analysis | Interpret images the conversation model cannot read directly |
| Image generation | Generate requested images |
| Audio analysis | Voice notes and audio files |
| Speech | Generate spoken replies |
Automatic chooses from configured providers. Optional roles can be disabled where the interface offers it. A missing image-generation model should not be diagnosed as a chat-model failure.
A conversation can override its model and supported reasoning effort. The gateway snapshots that choice for a run so it does not change underneath an in-progress turn. Internal workers can use tier-selected models; see collaboration.
Compatible and local model endpoints
Jian can call a separately operated OpenAI-compatible service. It does not install model weights or start that service for you. Verify the endpoint from the gateway's network, not only from your browser.
Local desktop mode permits loopback services, including plain HTTP. On a server, private destinations are governed by host-admin origin policy. In Docker, localhost is the container itself. See configuration before using a private address.
An endpoint advertising a familiar protocol does not guarantee support for every tool, image, streaming, or reasoning option. Select models appropriate to the task and check errors from the actual request.
Decisions and web search
Providers also contains optional service connections:
- Tavily: powers the agent's web search when configured.
- Decisions / Jev: provides second opinions on bounded questions such as memory usefulness, skill relevance, learning triage, and workflow decisions. It receives the specific message, action, or shortlist being judged.
Without a Decisions key, or when a particular use is disabled/unavailable, the caller uses its fixed rules. The switches select whether Jev participates in those decisions; they do not necessarily disable the underlying feature. In particular, the learning switch does not stop learning reviews.
Secrets, usage, and failures
Credentials entered in the panel are kept in the gateway's encrypted vault. Supported environment-backed credentials are detected from the server environment. They are not readable as ordinary model context.
Cloud requests still send the content needed for the task to the selected provider. Usage from conversation steps, compaction, media analysis, and learning can all contribute to cost. Treat displayed estimates separately from the provider's final bill.
For failures, separate invalid credentials, model availability, account limits, endpoint reachability, and an unsupported capability. Reconnecting a provider is not the first fix for every runtime error.
Implementation reference
Provider contracts, model default UI, and Decisions service.