Am-i-openai-compatible

am-i-openai-compatible (CLI: aioc) is a no-nonsense probe for any HTTP server that claims to speak the OpenAI API — plus a living catalog of how the OSS implementations actually behave.

Stack: Python ≥ 3.10. License: MIT. Public repo.

The term "OpenAI-compatible" gets used so loosely it's nearly worthless. llama.cpp, vLLM, Ollama, LM Studio, TabbyAPI, SGLang, MLC, comfy-openai — each picks its own subset of the surface and adds its own quirks. Some return usage, some don't. Some implement /v1/responses, most don't. Some honor seed, some pretend to.

What it gives you

• aioc probe URL — one command that tells you which OpenAI endpoints a server actually implements, which contracts it honors, and where it deviates. JSON report; renders as a table via aioc render.
• The compatibility docs — an opinionated reference for the real OpenAI surface and an honest matrix of how the major OSS servers compare. PRs welcome when behavior changes.

HT-compat profile

aioc probe URL --profile ht adds endpoints for model classes OpenAI doesn't pin yet: promptable segmentation (SAM3), omni-modal chat, reranking, layered image generation, audio extraction.

HT-compat v1.1 extends this to encoder-only BERT-style tasks: /v1/qa, /v1/ner, /v1/classifications — cherry-picked from the HF transformers pipelines and TEI's /predict semantics, wrapped in the OpenAI-style envelope already used by /v1/reranking. See docs/spec/ht-compat.md.

Sources

• README
• Docs.

See also

• legal-engine — consumes the HT-compat encoder endpoints.
• inventory — consumes reranking + segmentation.