One request, four API specs

A system prompt, one user message, temperature 0.2, a 256-token cap. Sent to OpenAI, Anthropic, Gemini and Ollama, that conversation is four different documents. moonmuse gives MoonBit one contract above all four, and reads and writes each of them with the same code.

4wire formats
1request type
17packages
0vendor names in code

Read the guide Edit all four

POST /v1/chat/completions

{
  "model": "gpt-4o-mini",
  "messages": [
    {"role": "system", "content": "Answer in one sentence."},
    {"role": "user", "content": "Why is the sky blue?"}
  ],
  "temperature": 0.2,
  "max_tokens": 256
}

POST /v1/messages

{
  "model": "gpt-4o-mini",
  "max_tokens": 256,
  "system": "Answer in one sentence.",
  "messages": [
    {"role": "user", "content": [
      {"type": "text", "text": "Why is the sky blue?"}
    ]}
  ],
  "temperature": 0.2
}

POST /v1beta/models/gpt-4o-mini:generateContent

{
  "systemInstruction": {
    "parts": [{"text": "Answer in one sentence."}]
  },
  "contents": [
    {"role": "user", "parts": [{"text": "Why is the sky blue?"}]}
  ],
  "generationConfig": {
    "temperature": 0.2,
    "maxOutputTokens": 256
  }
}

POST /api/chat

{
  "model": "gpt-4o-mini",
  "messages": [
    {"role": "system", "content": "Answer in one sentence."},
    {"role": "user", "content": "Why is the sky blue?"}
  ],
  "stream": false,
  "options": {
    "temperature": 0.2,
    "num_predict": 256
  }
}

The bodies in these tabs are written by this library, compiled to JavaScript and running on this page.

01 · The idea

Name the vendor once

A provider is anything that can answer a request. Which vendor, whether it runs in a data centre or on the laptop, and which JSON it speaks are details behind that one value.

MoonBit

let provider : &@moonmuse.Provider =
  @openai.Client::new(api_key="sk-…")

let request = @moonmuse.Request::new(
  "gpt-4o-mini",
  [@moonmuse.Message::user("Why is the sky blue?")],
  temperature=0.2,
  max_tokens=256,
)

let reply = provider.chat(request)
println(reply.text())

Provider

Ask for an answer, a stream, vectors, or the model list. Six implementations ship here, and a layer is a provider too — so retry, fallback, routing and caching wrap in any order.

Dialect

One vendor's wire format, read and written. The client and the gateway use the same value, which is why they cannot disagree about a field.

02 · What it does

Everything around one call

The request is the small part. These are the parts a second integration usually has to write again.

Streaming

One callback

Text, a tool call starting, its arguments arriving in pieces, then done — the same four deltas from a JSON chunk, seven named events, a candidate array or a line of NDJSON.

Tools

A schema and a function

Toolbox::call is the only place a call becomes a result, and a tool that fails produces an error result rather than an exception.

Layers

Retry, fall back, cache

Only failures another attempt could avoid are retried. Fallback switches on the same classification; the cache key covers every field that decides an answer.

Structure

An answer that fits

Ask for JSON against a schema, check the reply, and send it back with the reason when it does not fit — up to a budget.

Serving

Four APIs out of one

Point serve at any provider and it answers as OpenAI, Anthropic, Gemini or Ollama. The official client libraries cannot tell it from the vendor.

Observation

Traces under the conventions

One client span per call and token metrics, under the OpenTelemetry GenAI names. Message bodies stay out unless asked for.

03 · The module

One package per concern

A package that is not imported is not compiled or linked, so the local-model and serving halves cost nothing to a program that only asks for a completion.

moonmuseMessages, requests, replies, the two contracts, the event-stream reader

openaiThe chat format and every endpoint that speaks it

anthropicThe Messages API, its event stream, its token counter

geminigenerateContent, streamGenerateContent, embeddings

ollamaThe native /api endpoints and the newline-delimited stream

bedrockSigV4 signing and the Converse API, event-stream framing included

layersRetry, fallback, rate limiting, caching, routing

toolboxTool execution and the tool loop

structuredA JSON answer that fits a schema, with a repair loop

promptPrompt templates, rendered against a value map

modelsContext windows, output limits, prices, capabilities

countToken counting over a vocabulary

serveAny provider exported as four APIs

transportThe one place a call becomes bytes on a socket

localA local engine behind the same contract

otelTraces and metrics by the GenAI conventions

fakeA scripted provider, for tests above the wire

04 · Where to go