# Application inference

Use application inference when a feature needs a model response, such as summarizing a support ticket. Choose the provider for the application in **Services → Inference**. Use the managed route where available, or select an eligible connected provider for that environment. The model selected in the Project Composer controls the conversation.

## Try a one-sentence summary

From a hosted backend with application inference credentials, use the runtime’s `PLAYBOOKS_INFERENCE_URL` and `PLAYBOOKS_INFERENCE_KEY`. On a released Site backend, request `GET /models?capability=responses` with the key as a Bearer credential. Set `MODEL_ID` to an ID from that list. In a Project sandbox, use a response-capable model enabled for the application. Model discovery requires a Site runtime credential.

```sh
curl --fail-with-body "$PLAYBOOKS_INFERENCE_URL/responses" \
  -H "Authorization: Bearer $PLAYBOOKS_INFERENCE_KEY" \
  -H 'Content-Type: application/json' \
  --data "$(jq -n --arg model "$MODEL_ID" '{model: $model, input: "Summarize in one sentence: My invoice shows the annual plan but I selected monthly."}')"
```

Look for a generated summary that preserves the annual-versus-monthly discrepancy. If access fails, check the provider assignment, model access, and funding for that runtime. Keep the request in backend code so the visitor never receives the runtime key.