Bringing more control over your connectors

In a June 24, 2026 blog post, Mistral AI noted that making AI agents connect to enterprise data in demos is easy, but running them in production is difficult. The core issue: connectors must simultaneously obey the permission settings of the source platform and the control rules set by administrators in Mistral Studio or Vibe. Automated work must run on behalf of a user or service account, never impersonating the author. When a connection breaks, it must be possible to determine why. Administrators need to be able to decide, down to each tool, which parts of the organization can use which features.

These requirements are not new, but Mistral's latest update attempts to address them systematically. The blog post emphasizes that production-grade connections have 'several non-negotiables,' and the new features are designed around them.

The new admin console (GA) operates on two levels: team level and within the connector. Workspace and organization controls allow assigning independent connector access to each team. For example, the finance workspace can only access internal data sources with no open web access, while the engineering team gets developer tools and internet access. Different teams under the same directory can have different rules.

Tool-level control goes further: within any connector, individual tools can be turned on or off for the entire organization or a single workspace. You can block all write operations, or just the delete_file action on a knowledge base. The connector stays connected, but only approved tools can run. This granular control means administrators can finely manage risk without disconnecting the entire connector.

API keys now support connector scoping (GA). When creating a key, you can choose whether it can only access the workspace's shared connectors or also private connectors. This gives automated jobs only the minimum permissions they need. Combined with service accounts, automated work runs as a defined identity, not the author's identity. This directly addresses the 'identity impersonation' problem—a common security risk.

Multi-account connectors (GA) allow a single connector to bind multiple login accounts. For example, you can connect a personal account and a work account to the same connector, set a default account, and switch by task. Each account stores and refreshes tokens independently, so the agent can act under the correct account without deploying a second connector. This is especially useful for scenarios requiring access to different datasets, such as personal calendar vs. company calendar.

The connector debugger (public preview) aims to provide end-to-end root cause analysis. Users point to an MCP server URL, optionally add OAuth credentials, and run a check. The debugger steps through the connection, from reaching the server to opening an MCP session, in 11 steps, logging each one. The blog post notes that previously, issues like OAuth token exchange failures were guesswork; now the debugger shows exactly which step failed. This significantly reduces the difficulty of integration maintenance.

Connectors are now integrated into Workflows (public preview) and Vibe Code (GA). In Workflows, users can declare required connectors and accounts in the workflow definition, and Studio resolves these dependencies at runtime. Personal credentials are not exposed in automated work, and jobs do not fail midway due to token expiration. For example, a nightly task can pull data from CRM, update a tracker, and read a knowledge base across the entire job without disconnecting.

In Vibe Code, coding agents can type /connectors to choose from local MCP servers and admin-approved workspace connectors, enable needed tools, and start tasks. Asynchronous agents can access Outlook, Jira, Notion, Linear, and Confluence under the same rules. This ensures that connections in development environments are also governed.

Mistral says it is continuously adding connectors. The current catalog covers over 60 integrations, including Atlassian, BigQuery, Box, Google Drive, Notion, SharePoint, Databricks, Snowflake, Slack, Salesforce, Stripe, and more. When a desired platform is not listed, custom MCP connectors can fill the gap. The blog post does not mention specific timelines or future plans, but hints that the connector ecosystem will continue to expand.