Mistral Studio adds managed agents and one-command app hosting for Le Chonk

Mistral AI has added managed agents and apps to Studio, its enterprise platform, along with new workflow, connector, observability and evaluation features, the company said in a post on X. The pitch: Le Chonk supplies the model, Studio runs the work.

The company’s announcement post describes six pieces. The post does not give prices, quotas or a rollout schedule beyond “today”, and points readers to its pricing page.

What Studio now hosts and traces

Managed agents are multi-step agents that can use skills (team procedures), sandboxes for running code, subagents for delegated tasks and hooks for custom logic. Studio operates the runtime, while developers set limits on the tools, data and actions an agent can reach. Mistral says the runtime sits on infrastructure that handles millions of Mistral Vibe sessions each week.

Workflows let teams decide which steps follow fixed rules, where an agent reasons and where a person approves. Interrupted runs can resume. New are event triggers from enterprise systems and a visual graph of the logic. Workflows can also be generated, tested and deployed from Mistral Vibe as editable code.

Apps come with hosting, authentication and access controls, plus pre-approved reusable components meant to shrink the amount of new code that needs a security review. They can use agents, workflows and search indexes, and deploy with one command to their own link or inside Vibe.

For data access, models connect through connectors and MCP servers. A new Mistral MCP server lets developers build locally and deploy to Studio with one command. Observability uses OpenTelemetry traces that show each run step, including model calls, tool use and approvals. Evaluations let teams test changes to prompts, models or tools against their own datasets before release.

Who can run it, and where

Mistral says Studio and its models can run in a customer’s own cloud, VPC or on-premises environment, on Mistral-hosted deployments, or a mix. It names no specific regions. ABANCA, Accenture, Alten, BNP Paribas, Capgemini, CMA CGM and Tata Consultancy Services are listed as using Studio in production.

The post cites two customer figures: an incident investigation agent that saves an engineering team “hundreds of hours each month”, and a product operations triage workflow that saves about 35 hours a week. Both are Mistral’s claims, with no named deployment or method, and have not been independently verified.

A model company moves up the stack

The release fits a pattern in Mistral’s recent output. On October 29 the company released the weights of Le Chonk, its Mistral Large 4 model. According to our earlier reporting, the model carries a 19.7% score on FrontierCode 2.0, the top open-weight result in Mistral’s own chart, and costs $0.70 per million input tokens and $2.10 per million output tokens through the API. Those prices matter here because agents burn tokens: a hosted runtime that loops through tools multiplies the per-token bill, and Studio is where that usage would be metered.

Studio also ties into the consumer-facing product. Earlier this month Mistral’s assistant became Vibe, and the new apps can be deployed inside it, so one workspace now serves as both the assistant and the place internal tools land. The company has also been courting engineering users, announcing open benchmarks with Nvidia for agents on chip and CAD work.

Open weights give Mistral a different argument from the large US labs: customers can run the same model on their own hardware, and Studio can follow it there. That is the sovereignty case for European buyers such as banks, and several of the named customers fit it.

What the announcement leaves open

Several details are missing. There are no prices for managed agents or hosted apps, no stated limits on sandbox time or concurrent runs, and no indication of which features are available outside Mistral-hosted deployments. The claim that the pre-approved components cut security review effort is not backed by a measurement. Nor does the post say whether Studio agents are restricted to Mistral models or can call others.

Developers can start from the Studio page and the pricing page Mistral links. Teams weighing it against agent platforms from other labs will need those missing numbers before comparing costs.