Mistral AI Connectors release notes / what's new

Mistral AI has introduced Connectors in Studio, allowing developers to build AI applications grounded in enterprise data using both built-in and custom Model Context Protocol (MCP) connectors. These connectors are now available via API and SDK, making them accessible for all model and agent calls.

Centralized Integration Management via MCP

Mistral AI Connectors package integrations into reusable entities using the MCP protocol. This approach removes the need for developers to manually write and maintain tool functions, handle OAuth token refreshes, or manage API documentation for every new agent.

Once a connector is registered, it is centrally available across Mistral applications, including LeChat and AI Studio (with Vibe support coming soon). This centralization prevents the duplication of integration logic across different teams and reduces security risks associated with arbitrary code implementations.

Custom MCP Connector Implementation

Developers can create custom MCP connectors by specifying the server URL and authentication configuration. For example, a connector can be registered with a specific client_id, scopes, and redirect_uri for OAuth integration with a CRM like Salesforce.

Once registered, attaching a connector to a conversation requires only a single line of code:

response = client.beta.conversations.start_async(
    model="mistral-medium-latest",
    inputs="Which enterprise accounts renewed last quarter?",
    tools=[{"type": "connector", "connector_id": "salesforce-crm"}],
)

Advanced Tool Control and Governance

Mistral AI has introduced several features to provide developers with precise control over how tools are invoked and executed.

Direct Tool Calling

For deterministic workflows where the model does not need to decide when to invoke a tool, developers can now call connectors directly. This is useful for debugging and pipeline-style automation to limit ambiguity.

result = await client.beta.connectors.call_tool_async(
        connector_id="my_deepwiki",
        tool_name="read_wiki_structure",
        arguments={"repoName": "sqlite/sqlite"},
    )

Human-in-the-Loop Approval

To ensure security and governance, developers can implement human-in-the-loop approval flows. By using the requires_confirmation parameter in the tool_configuration, the system pauses execution and returns control to the application, requiring explicit user confirmation before a tool is executed.

Granular Tool Configuration

Developers can control tool availability within a connector without modifying the connector itself. The tool_configuration setting allows users to exclude specific potentially damaging actions (e.g., excluding delete_file from a GitHub connector) to maintain safety.

Programmatic Access and Ecosystem

Connectors are available in Public Preview and can be managed programmatically through the API/SDK for the following actions:

• Creating, modifying, listing, and deleting connectors.
• Listing the tools associated with a connector.
• Directly running tools.

These capabilities are accessible via the Conversation API, Completions API, and Agent SDK, facilitating integration with enterprise systems such as knowledge bases, CRMs, and productivity tools.