Mistral AI gives admins fine-grained control over enterprise connectors

Mistral AI has updated Connectors, its integration system for enterprise AI agents. The company rolled out six new features at once, aimed at making agent connections to real corporate data work not only in demos but in production.

Granular control for administrators

When an AI agent runs in an enterprise environment, it faces two sets of restrictions: those defined in the source system (a CRM, a repository, a knowledge base) and those the administrator configured on the Mistral platform. Until now, the two layers fit together poorly. The new Enriched Admin Controls close that gap.

Workspace-level management lets each team have its own set of connectors. The finance department works only with internal data sources and has no internet access; engineers get developer tools and the open web. One corporate user directory, different rules for each team, and no need to create separate environments.

Tool-level management lets you disable a specific action inside a connector. For example, you can forbid delete_file in a knowledge base while keeping search and read access. The connector stays connected, but only the tools the administrator has explicitly allowed will run.

API keys with a restricted scope address the identity problem in automated tasks. When you create a key, you choose whether it reaches only the shared workspace connectors or the users' private ones as well. Paired with service accounts, an automated task runs as a specific service entity rather than impersonating the person who wrote the script.

Multiple accounts and failure diagnostics

Multi-Account Connectors let a single connector hold several accounts. A personal and a work Google account can be attached to the same connector, with one active by default, and the agent switches between them depending on the task. Each account is stored and refreshed independently, so you don't need to stand up a separate connector for every scenario.

Connectors Debugger is a diagnostic tool for MCP connections, currently in Public Preview. You enter the MCP server URL, add OAuth credentials if needed, and the debugger walks through 11 steps in sequence, from establishing a TCP connection to the server to opening an MCP session. Each step is logged separately.

Previously, a broken connection meant hours of guesswork: is it OAuth? The network? The server? The token? Now the debugger points to the specific step where the failure occurred. Support shifts from "look everywhere" to "look here."

Connectors in Workflows and Vibe Code

Two new points of use extend Connectors beyond interactive AI chats:

• Connectors in Workflows (Public Preview): connectors are declared directly in the Workflow definition. Studio resolves dependencies at launch. Long-running and recurring tasks are no longer interrupted by authentication expiring midway through execution.
• Connectors in Vibe Code (GA): the same connectors already configured in Studio are available in the developer interface. No reconfiguration is needed: reuse the integrations you've built without extra steps.

There is a single point of control: configure once, use in chat, Workflows and the development environment. In enterprise settings this is critical, because duplicated access settings quickly produce inconsistencies and vulnerabilities.

What this means

Connecting an AI agent to corporate data in a demo is easy: a couple of keys and a wrapper are enough. Production demands a different level: manageability, auditing and predictable behavior during failures. Mistral is systematically addressing these requirements: tool-level permissions, transparent identity for automation, built-in diagnostics. Enterprise agents are moving from the "works in the console" phase to the "works safely inside the organization" phase.