Examples

Small, runnable projects that call an OpenAI-compatible gateway for text, images and video. They are written to be copied into real code: real error handling, real timeouts, real polling — not three-line toy snippets.

Project	Stack	Run
ts-media	TypeScript, openai SDK, tsx	npm i && npm run image -- "a corgi on the moon"
py-media	Python 3.9+, openai + requests	pip install -r requirements.txt && python -m apimaster_media.image "a corgi on the moon"
open-webui	Open WebUI function (Python)	Paste into Admin Panel → Functions, set the key in the valves

Both read APIMASTER_API_KEY and APIMASTER_BASE_URL from the environment, so the same code runs against any other gateway by changing one variable.

What they demonstrate

• Streaming chat with a measured time-to-first-token, and multi-turn done the way that actually works on an aggregator: the full history is resent every turn.
• Image generation, sync and async. Sync is simpler; 2k/4k jobs can exceed the gateway timeout and return 408, and the code tells you to switch to async when that happens instead of failing cryptically.
• Video generation with polling and download, including the aspect-ratio trap that makes portrait image-to-video come back letterboxed.
• Retry policy that matches reality: 429 and 5xx back off; 400/401/402 fail fast.

The four rules these examples encode

1. Send the whole conversation history every turn — previous_response_id and store do not hold context on this gateway.
2. Use /chat/completions for Claude model ids; /responses returns 500 for them.
3. When parsing /responses, find the output item with type == "message" rather than indexing output[0] — a reasoning item can come first.
4. Never hardcode a model id you have not listed with GET /v1/models.