Mistral's connectors move agent permissions into config

Mistral's case for Connectors in Studio, released in public preview on May 22, 2026, starts from a familiar complaint. Teams inside the same company keep rebuilding the same integrations, each with its own OAuth handling, token refresh, and no shared view of the traffic. Mistral's answer is to wrap each integration as a reusable MCP server, register it once in the platform, and attach it to any conversation or agent in one line. Built-in connectors for GitHub, Gmail, and web search sit alongside custom ones.

A tool_configuration block lets you include or exclude specific tools from a connector, and requires_confirmation pauses execution so your application decides whether a proposed tool call runs. That puts the line between model judgment and human judgment in code you can review. This matters much more for agents doing real work than benchmarks. One thing I noticed, however, is that Mistral's own GitHub example excludes delete_file and allows everything else. In production I would much rather be forced to whitelist required tools and deny everything else by default.

The portability question is worth examining before standardizing on this approach. MCP is an open protocol but the registry, the OAuth configuration, and the approval flow would live in Mistral's platform. For a European team that picked Mistral partly to reduce dependence on US clouds, that may be a reasonable trade-off, but it is still a trade-off. Moving to another vendor would mean rebuilding that layer.