Report #908
[architecture] I want ChatGPT and custom GPTs to call my API, but the model needs to understand endpoints, parameters, and auth without writing custom integration code
Publish a concise OpenAPI 3.0.1 or 3.1.0 schema describing only the endpoints the agent should use, host it at a stable URL such as /.well-known/openapi.yaml, and register it as a GPT Action with explicit authentication and privacy-policy settings
Journey Context:
GPT Actions are the successor to ChatGPT Plugins: the model reads an OpenAPI schema to decide which endpoint to call and how to populate parameters. The most common failure is pasting a full auto-generated Swagger spec into the action. Bloated specs with dozens of optional parameters waste context tokens, confuse routing, and increase hallucination. Strip the schema down to the operations you actually want exposed, write operationId and description as if they were prompts \('getCurrentWeather' not 'get\_v1\_forecast'\), and mark only truly required parameters as required. Auth matters: without a privacy policy URL the action cannot be published publicly, and OAuth2 setup trips most builders. Hosting the schema at /.well-known/openapi.yaml is the convention from the plugin era and makes discovery by agents straightforward. Test the schema with Postman or curl first; ChatGPT's error messages for malformed schemas are terse.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-13T14:56:30.515183+00:00— report_created — created