Introduction

MMSP, the Model Message Stream Protocol, is one message format and one streaming grammar for every model provider. It ships as a Python package and a TypeScript package that behave the same way.

Model APIs agree on the building blocks: messages, thinking, tool calls, streaming. They disagree on the encoding. Anthropic streams content blocks, OpenAI streams output items, Gemini streams steps, and Chat Completions streams chunks that name nothing. MMSP reads each of them on its own terms and hands your code one shape.

Three objects

Everything you send and receive is one of three plain objects:

Object	What it is
UniConfig	The options of a request: tools, thinking level, system prompt, and the rest.
UniMessage	One message of the conversation, from the user or the assistant. It holds complete content items.
UniEvent	One event of a streamed response. A stream is delta events closed by one stop event.

They are dictionaries in Python and typed objects in TypeScript, with the same field names in both.

How a request travels

1. You create an AutoLLMClient with a model id and a client type: the vendor's official client, or a compatible client for any endpoint that speaks its protocol. A model id of a known family names its official client on its own.
2. You call a streaming method with messages and a config. The client translates both into the provider's request.
3. The provider streams its own events. The client turns each one into UniEvent fragments, and MMSP closes every item with its complete form.
4. The complete items, in order, are the assistant message. The stateful call records it in the client's history before it yields the last event.

Whatever the provider sent that your code does not need but the provider wants back, such as a thinking signature, travels with the item as fidelity.

Stateful or stateless

streaming_response_stateful takes the new message and keeps the history in the client. It is the one to use for agent loops: thinking and tool calls of earlier turns go back to the model exactly as they arrived.

streaming_response takes the whole message list every time and keeps nothing. Use it when you store the conversation yourself.

Where to go next

• Quickstart streams a first response and runs a tool call.
• The agent loop is a complete agent in forty lines.
• Streaming has the grammar every stream follows.
• Models and endpoints lists what MMSP talks to.