Build agents

mistral.rs can run the tool loop for you, expose standard OpenAI tool calls to your client, or act as the local runtime behind an agent app. Unlike a plain OpenAI-compatible model server, mistral.rs can execute tools locally and stream model text, tool progress, files, media, and session state from the same request. Tutorial 5 is the end-to-end walkthrough.

The agent system has three layers:

Tool protocol: how a model requests external work.
Tool execution: built-in code execution, web search, MCP tools, callbacks, or external dispatch.
App runtime: streaming model output, tool progress, files, generated media, and sessions as one local surface.

Choose a path

Learn the basics

Tool calling basics: OpenAI tool calls and the mistral.rs server-side loop.
Strict tool calling: constrain tool arguments to the declared JSON Schema.

Use built-in tools

Enable code execution: Python execution, file outputs, media capture, and sandbox isolation.
Web search: built-in search, extraction, and embedding reranking.

Build an app runtime

Agentic runtime for apps: streaming model output, tool progress, files, generated media, and sessions.
Persist agent sessions: state across HTTP requests.

Bring or expose tools

Connect to an MCP server: use external MCP tools.
Configure the tool loop: round caps and dispatch URLs.
Expose mistralrs as an MCP server: let other agents call your models.

Recommended order

For design rationale (server-side tool loops, session splicing), see the explanation section.