Mistral Adds Mini Apps, Sandboxed Code and Local Evaluation Workflows

Vibe Work can generate self-contained React mini apps from plain-language descriptions, render them immediately and place an interactive preview inline in the chat. [1] Users can continue the conversation to modify an app, edit its source files directly, or switch the Canvas panel between the running preview and a code view. [1] The complete source can be downloaded as a ZIP for use in an editor or deployment to a static host, while public links can expose only the app without sharing the associated chat messages. [1]

On paid plans, Code Interpreter executes Python and TypeScript directly within Vibe Work and replaces the legacy Chat code interpreter with the same capabilities. [4] Its native sandbox is isolated, has no internet access, cannot reach the user’s system or external services, and includes common data-analysis and charting packages. [4] Uploaded files are available for the current conversation only, and preinstalled Python packages include pandas, numpy and matplotlib. [4] Users can request analysis, charts, calculations and file transformations in plain language, receive tables, figures or files inline, and refine the output through follow-up prompts. [4]

Setting `local=True` runs an evaluation without sending data to Studio, removes the need to specify a project or evaluation, and leaves results available only in the terminal or as JSON. [2] Mistral recommends adjusting tasks and scorers locally before removing local mode and adding project and evaluation details to track results over time in Studio. [2] A single run can also receive any number of evaluators as a list, with the SDK executing each evaluator and calculating its statistics independently so that every metric appears separately in Studio. [3]

The documented workflow separates three jobs that are often intertwined: shaping an interface with Mini Apps, executing data-oriented code in an isolated environment, and measuring model behavior through evaluation runs. Local mode provides a practical boundary between experimentation and persistent Studio tracking, while multiple evaluators let teams preserve separate measures rather than compressing requirements into one metric. Code Interpreter and Mini Apps are complementary rather than interchangeable: one returns computed tables, figures or files, while the other produces an interactive React application that can be downloaded or shared. What the sources don't address: how generated apps and evaluation configurations should be reviewed, secured and governed before production deployment.