OpenAI-compatible

A ModelProvider speaking the OpenAI chat-completions wire format against a configurable base URL. Ollama, vLLM, llama.cpp, LM Studio, LiteLLM, OpenRouter, Azure and OpenAI itself all speak it.

That is what makes a local, air-gapped install a configuration choice rather than a fork.

No SDK

It depends only on the copilot’s domain package — for the port’s type, erased at runtime — and Node built-ins. The surface is one POST /chat/completions and a few lines of event-stream parsing, and a dependency for that would not earn its place.

Install

npm install @apograph/copilot-provider-openai

import { createOpenAiProvider } from '@apograph/copilot-provider-openai';

createOpenAiProvider(config.plugins.copilot.providers.openai);

Configuration

{ baseUrl, models, apiKey?, headers?, capabilities?, timeoutMs?, maxRetries?, maxTokensField? }

Pointing it somewhere

The model list is a list

One endpoint and one credential back several models, so a user can switch mid-conversation. A run naming a model outside the list raises an unknown-model error rather than quietly answering on the default.

Two pieces worth knowing about

The retry policy is pure. Which statuses retry, the backoff, and honouring Retry-After live in their own module with their own tests, and it runs pre-stream only — once bytes are flowing, a retry would replay a partial answer.

Tool-call fragments are assembled by index. Parallel tool calls interleave on the wire, arriving as fragments that have to be reassembled per call. That accumulator is the only piece here holding real state, so it is pinned directly rather than covered only through the provider.