The OpenAI-compatible API
CONNOR has an API that works like OpenAI’s chat completions API. Code and tools built for OpenAI’s API can ask CONNOR questions instead. Replies come with their sources, as they do in chat.
What you need
- A Chat API key. See API keys. An MCP key doesn’t work with this API.
- CONNOR’s Chat API base URL. The API keys page shows it under Where to point them, with a button to copy it.
Point your tool’s OpenAI base URL at CONNOR’s, and use your Chat API key as the API key.
Which model to ask for
Use the model your tool finds when it lists the models (the /v1/models endpoint). CONNOR serves one model id there. The API keys page doesn’t name a model, because CONNOR chooses which AI model answers.
What the API supports
- Chat completions, streaming or not.
- Listing the models.
How API requests are charged
Each request is a turn, charged from your credit like a turn in chat. It counts towards the key’s monthly spend limit, if it has one. If your credit or the key’s limit can’t cover it, the request is refused.
If your code stops waiting for a reply, CONNOR stops the turn. You pay for the work done up to then.
Usage shows what each key spent. See How credit works.
How API requests differ from chat
- Each request is its own conversation. Send the earlier messages with each request if you want CONNOR to follow a conversation.
- API conversations don’t go into CONNOR’s memory.
- Nothing can be approved over the API. Anything that needs your approval, such as starting an eval, happens in CONNOR.
- Each key can run only a few requests at once. Extra requests are refused with a rate-limit error until one finishes.
- Errors come back in OpenAI’s error format.