Mistral Launches Agents API with Code Execution, Web Search, and MCP Tool Integration

Mistral AI has released its Agents API, a production framework for building AI agents that can execute code, search the web, and coordinate multi-step workflows. The API launched on May 27, 2025.

Core Capabilities

The Agents API provides four built-in connectors:

Code execution: Agents can execute Python code in a sandboxed environment for mathematical calculations, data visualization, and scientific computing.

Web search: Integration with web search sources enables agents to access current information. According to Mistral, Mistral Large with web search scores 75% on the SimpleQA benchmark compared to 23% without web search. Mistral Medium achieves 82.32% with web search versus 22.08% without.

Image generation: Powered by Black Forest Lab's FLUX1.1 [pro] Ultra model for creating visual content.

Document library: Built-in RAG functionality allowing agents to access documents from Mistral Cloud.

Model Context Protocol Support

The API supports tools built on the Model Context Protocol (MCP), an open standard for connecting agents to external systems including APIs, databases, and documents. Mistral has published cookbooks demonstrating MCP integrations with GitHub, Linear, and financial data sources.

Stateful Conversations and Memory

The API maintains conversation context across interactions through persistent memory. Developers can create conversations in two ways: by specifying an agent_id for specialized capabilities, or by directly configuring model and completion parameters.

The system supports conversation branching, allowing developers to view conversation history and initiate new paths from any prior point. Streaming output is available for real-time interactions.

Multi-Agent Orchestration

The API enables dynamic orchestration where multiple agents collaborate on complex tasks. Agents can hand off tasks to other specialized agents—for example, a finance agent delegating to a web search agent or calculator agent based on context.

Mistral states this allows "a single request to trigger tasks across multiple agents, each handling specific parts of the request."

Pricing and Availability

Pricing details were not disclosed in the announcement. The API is available now through Mistral's platform.

What This Means

Mistral's Agents API directly competes with OpenAI's Assistants API and Anthropic's tool use capabilities by providing a complete framework rather than requiring developers to build orchestration logic themselves. The 52-percentage-point improvement in SimpleQA scores with web search (75% vs 23% for Mistral Large) demonstrates measurable impact of external tool integration on factual accuracy. MCP support positions Mistral to benefit from the growing ecosystem of standardized agent tools, though MCP adoption remains early-stage across the industry.