Model and API settings
Operators configure models, upstream APIs, the prompt assistant, storage, and notifications. Users can call only permitted models and presets; the browser never receives provider keys or upstream routing details.
Manage presets
Presets define the API path, default model, response format, and visible parameters. Before making a change, confirm that:
- The upstream address is within the allowed host range.
- The model is compatible with the path.
- The new preset has passed a health check.
- Existing users will not lose a configuration they are using because a preset is deleted or switched.
Current production upstream
JumpCLI's default production upstream uses an OpenAI-compatible Images API:
| Item | Verified value |
|---|---|
| Upstream address | https://www.xinyue.mom |
| API path | /v1/images/generations |
| Default model | gpt-image-2 |
| Allowed host | www.xinyue.mom |
gpt-image-2 has passed JumpCLI's complete preview-generation, image-decoding, and validation flow and can be used as the default model. Provider keys remain only in the server environment and must not be copied into preset names, descriptions, tickets, or browser requests.
The same upstream also provides grok-imagine-image, grok-imagine-image-2.0, and grok-imagine-image-lite. Grok image models must use quality=low or quality=medium; quality=auto is rejected by the upstream with a 400 response. This upstream may return non-HTTPS image links when response_format=url; JumpCLI rejects those downloads under its security policy. Production Grok presets should use quality=low (or medium) and response_format=b64_json. This combination has passed image-decoding and format validation. Before exposing it to users, confirm that UI defaults or caller requests always use these compatible parameters.
JumpCLI currently activates only one preset and does not automatically balance traffic among providers. When multiple providers must be switched according to cost, capacity, or failure state, deploy an independent OpenAI-compatible model gateway in front of JumpCLI and configure that gateway as the sole upstream address.
Health checks
The settings page can check connections to the upstream, object storage, prompt optimizer, and AI assistant. A successful health check proves only that the current configuration is reachable. Submit one minimal generation request with a test account to validate the complete flow. If a request returns billing_not_ready, first complete Meteroid billing projection and initialize the credit cycle for that organization. Do not bypass the billing gate by directly modifying the tenant database.
Key handling
Keys are supplied at service startup through a secret registry or environment variables. The UI can select only declared key references and never displays raw values. Record the change time, impact scope, and rollback plan for every key rotation.
For deployment-related variables, see Environment variables.