Connect the dots: Build with built-in and custom MCPs in Studio

Mistral AI released a public preview of Connectors for Studio, with the main theme of making it easier and more secure for enterprise AI applications to connect to internal or external data sources. In the past, when building agents or workflows, teams often had to repeatedly write various integration layers: looking up API documentation, maintaining tool functions, handling OAuth, token refresh, pagination, and error conditions. Mistral's approach is to encapsulate integrations into reusable connectors, registered as first-class resources within the platform via the MCP protocol, so the same connector can be reused by a conversation, agent, or workflow without rewriting the logic in different projects. This update supports using the API/SDK to create, modify, list, and delete connectors, as well as list the tools within a connector and execute them directly. Both built-in connectors and custom MCP can be used by the Conversation API, Completions API, and Agent SDK, and are centrally registered to the Mistral platform for use by LeChat and AI Studio, with Vibe also expected to support them. The article demonstrates a combination of a Salesforce CRM connector, DeepWiki MCP, GitHub, and web search, used to build an agent that can analyze repositories, documents, and real-time web data. For developers, direct tool calling is another important capability, because not all processes are suited to letting the model decide when to call a tool; directly calling a connector tool gives better control over debugging, pipeline-style automation, and more deterministic processes. On governance, Mistral added human-in-the-loop approval, which can be configured so that certain tools require confirmation before execution—such as sensitive actions like Gmail search—letting the model first propose a tool call, then having the application or user decide whether to proceed. Overall, this is a step by Mistral to strengthen enterprise AI agent infrastructure, with the focus not on new model capabilities but on reducing the cost of enterprise data integration, permission control, observability, and duplicated development.