#Servers

The client sends every request to its base URL plus the endpoint's path (/chat/completions, /embeddings, /models), so any server that serves those paths with OpenAI's fields works. The base URL is the part before /chat/completions.

def router as openai.Client init openai.withBaseUrl(openai.client(os.getEnv("OPENROUTER_API_KEY")),
    "https://openrouter.ai/api/v1");
$router = openai.withHeader($router, "X-Title", "my tool");

withHeader adds a header to every request (OpenRouter's HTTP-Referer and X-Title, a gateway's own); the headers the client sets itself are refused. A server behind a private CA takes its certificate with openai.withTls($c, http.TlsOptions{caCert: fs.readBytes("ca.pem")}).

Each server implements its own subset: a field one does not know is ignored or refused with a 400, a model name is the server's own (qwen2.5:0.5b on Ollama, meta-llama/Llama-3.1-8B-Instruct on vLLM). models lists what it offers.

#What the deck reads from any server

• An error body in OpenAI's shape, {"error": {"message", "type", "param", "code"}}, with code a string or a number; a bare {"error": "text"}; the error fields at the top level ({"object": "error", "message", "type", "code"}); FastAPI's {"detail": "text"}; and a plain text body (Ollama answers an unknown path with 404 page not found as text/plain). An HTML page gives only the status text.
• A 2xx reply holding an error object is an error, as some proxies send them.
• Tool-call content given as "" instead of null, an extra index on each call, missing usage details: read as OpenAI's. addReply drops the fields a request does not carry.
• Message content as a list of parts (Mistral's reasoning models): text joins the text parts.

#Ollama

Checked live with Ollama 0.40.16 (ollama/ollama in Docker, bound to 127.0.0.1) and the models qwen2.5:0.5b and all-minilm: chat with the developer role and max_completion_tokens, two parallel tool calls in one reply and the answer after their results, a no-argument tool (its arguments "{}"), structured output with a strict json_schema, embeddings as floats and as base64, and /v1/models. Errors come in OpenAI's shape (not_found_error for an unknown model, with the HTTP status), except for an unknown path. Its usage reports prompt_tokens_details.cached_tokens. Ollama's native API (/api/chat, /api/embed) is different and not this deck's; call reaches it only below the base URL, so use a second client with the base URL http://127.0.0.1:11434 for it.

#Azure OpenAI

Two schemes, both with the key in an api-key header:

• v1 (azure): https://<resource>.openai.azure.com/openai/v1/... (or .services.ai.azure.com), the same paths as OpenAI, no api-version since the v1 API went GA in August 2025. The request's model is the deployment name. This is the one to use.
• Dated (azureVersioned): the deployment in the path, /openai/deployments/<deployment>/chat/completions?api-version=..., with AZURE_API_VERSION (2024-10-21, the last dated GA version) or a preview. The deck takes the deployment from the request's model, which also stays in the body; models lists /openai/models, and call puts /openai before its path and adds the api-version.

A deployment name goes into the path, so it is checked: 1 to 64 letters, digits, ., _ and -. Azure's request id is apim-request-id; OpenaiError.requestId holds it.

#Any other endpoint

call(c, method, path, body) sends any request below the base URL with the client's key, headers, deadline, cap, retries and errors, and returns the JSON reply (JSON null for an empty body). body is JSON null for none.

def m as json.Value init openai.call($c, "GET", "/models/gpt-5.4-mini", json.decode("null"));
def mod as json.Value init openai.call($c, "POST", "/moderations",
    json.map({"model": "omni-moderation-latest", "input": "some text"}));

The path starts with one / and may carry a query; . and .. segments, spaces, # and backslashes are refused, so a path cannot leave the base URL.