The whole point is the handoff: when a problem calls for a different model's take, the question goes out and the answer comes back inline, so the surrounding context and your notes stay in one place. Keep expectations proportionate to the surface — one tool, a message list and a model name, with no streaming, no tool use and no conversation state on the other side.
A one-tool bridge to OpenAI's chat completion API. You pass the messages and name the model; the answer comes back into the conversation you were already having.
- openai_chat sends a message list to OpenAI's chat completion API and returns the reply
- The model is an argument — gpt-4o, gpt-4o-mini, o1-preview and o1-mini are the ones the project lists
- A message array rather than a single string, so prior turns can travel with the question
- Errors surface rather than disappearing, which is all the handling the project claims
An OpenAI API key from the OpenAI platform, set as OPENAI_API_KEY in the client's env block; calls are billed against it. Node.js 18 or higher, started as npx -y @mzxrai/mcp-openai@latest.
One command plus a key — npx -y @mzxrai/mcp-openai@latest, then supply credentials