Mistral's switch to MCP connectors leaves data governance in limbo

Mistral is migrating away from its own indexed document retrieval system to the Model Context Protocol (MCP), a move that leaves critical governance questions unanswered for enterprise customers using Vibe Work to access company documents stored in Google Drive or SharePoint.

The current setup: indexed retrieval#

Today, administrators pre-select which Google Drive folders or SharePoint sites should be searchable. Mistral then processes and indexes those files, storing the index in European data centers. Users authenticate with their personal accounts, and when they search, the system cross-references permissions from the original source to ensure they see only what they can access. Synchronization tasks automatically incorporate deletions and changes—a process that can take anywhere from minutes to hours depending on data volume, but completes before users start searching.

MCP fundamentally changes the architecture#

The company is pivoting to MCP—a protocol that allows language models to call external tools and retrieve data directly from services. Rather than maintaining its own indexed cache, Mistral will hand off retrieval to third-party MCP servers that it does not operate. The announcement offers scant detail on whether those servers query APIs live, maintain separate search indexes, or use some hybrid approach.

This flexibility is MCP's strength in some contexts, but it means each connector's behavior depends entirely on whoever runs it. Mistral cannot guarantee data handling, performance, or retrieval quality under the new system.

Governance gaps left unaddressed#

The existing Google Drive connector respects file-level sharing rules—including group and domain access—so files marked "anyone with the link" are not automatically visible to the entire organization. SharePoint integrations rely on Microsoft Entra ID groups. The migration notice makes no commitment that the MCP versions will enforce the same permission boundaries.

While OAuth can limit what a server accesses on behalf of a user, it doesn't automatically translate to the same fine-grained search result filtering that Mistral's prebuilt index provided. The protocol simply wasn't designed for that kind of enterprise permission enforcement; the connector operator must build it in.

Additionally, Mistral offers no clarity on whether customer data will be cached, where it might reside, or how deletions will be handled after the August shutdown. Administrators remain uncertain whether they should manually disconnect old connectors before the deadline or whether Mistral will trigger deletion automatically.