Studio

Mistral Studio is one place to build AI agents, workflows, and apps, run them in production, and improve them on real usage. You compose them from shared components, connect them to your company's data and tools, and deploy them without assembling the infrastructure yourself. Once a system runs, you can trace what it does and evaluate each change on real cases before you ship the next version (Enterprise-tier organizations).

Agents and workflows run on Mistral's cloud or on your own infrastructure, and apps run on Mistral Cloud. Secrets, permissions, and guardrails are built in.

This section is for developers: you build, deploy, and run through code you control, with the SDKs, the API, the Mistral CLI, and the coding agent you already use.

Agents, workflows, and API requests authenticate with an API key from your Studio workspace.

1. Open the Studio console and go to API Keys (direct link).
2. Select New key, give the key a name, and set an expiration date.
3. Copy the key. It appears only once.
4. Store the key in the MISTRAL_API_KEY environment variable, where the SDKs and the examples read it:

export MISTRAL_API_KEY="your-api-key"

export MISTRAL_API_KEY="your-api-key"

The Agents SDK and the Workflows SDK are Python packages, and both need Python 3.12, 3.13, or 3.14 (>=3.12,<3.15). Create a project with uv, and add both SDKs to it:

uv init --python 3.12 my-project
cd my-project
uv add mistralai-agents mistralai-workflows

uv init --python 3.12 my-project
cd my-project
uv add mistralai-agents mistralai-workflows

import asyncio

from mistralai.agents import agents

agent = agents.Agent(
    name="hello-agent",
    model="mistral-medium-latest",
    instructions="You are a friendly assistant. Answer in one sentence.",
)

async def main() -> None:
    state = await agent.run("What can you help me with?")
    match state:
        case agents.Completed(output=answer):
            print(answer)
        case agents.Failed(error=error):
            print("The run failed:", error.message)

asyncio.run(main())

import asyncio

from mistralai.agents import agents

agent = agents.Agent(
    name="hello-agent",
    model="mistral-medium-latest",
    instructions="You are a friendly assistant. Answer in one sentence.",
)

async def main() -> None:
    state = await agent.run("What can you help me with?")
    match state:
        case agents.Completed(output=answer):
            print(answer)
        case agents.Failed(error=error):
            print("The run failed:", error.message)

asyncio.run(main())

I can help you with a wide range of tasks, including writing code, analyzing data, managing files, automating workflows, searching for information, and integrating with external tools or services.

I can help you with a wide range of tasks, including writing code, analyzing data, managing files, automating workflows, searching for information, and integrating with external tools or services.

Studio builds three kinds of systems. Choose by the shape of the process:

A process can combine all three. People can also use your agents and workflows in Vibe before you build an app for them. To compare the three in detail, see Choose what to build.

Domain experts can describe a process in plain language and get a first version of the workflow, which developers then refine in code: see Build in Vibe.

Every system you build needs the same production services. Studio provides them once, and each project reuses them:

• Runtime: deployments run agents and workflows on your machine while you build, on your own infrastructure, or managed by Mistral. Apps run on Mistral Cloud.
• Company context: Connectors reach your tools and data, Libraries and Agentic Search supply your documents, and prompts and skills keep versioned instructions and know-how that agents and workflows reuse.
• Controls: Secrets Manager stores credentials, governance sets who can use which data, tools, and models, and moderation and guardrails check what goes in and out of a model.
• Traces and evaluations: observability shows model calls, tool calls, agent activity, and workflow runs in one view, and evaluations compare versions on real cases before a release. Both are available to Enterprise-tier organizations only.

Once an admin connects an app for the organization or shares a connector with the workspace, the next project can reuse it: an agent, a workflow, and an app can call the same connector without rebuilding the integration.

Studio is one part of the Mistral platform. If you need something else, start from the matching section:

The Studio console shows what runs, such as executions, traces, and evaluations, and lets you create some resources without code. You can build in the console, in code, or with a coding agent, and switch between them.

Studio is organized around a recommended loop: build a system, give it context, test it, deploy it, observe it in production, and improve it, then start the next version from Build.

Read the loop clockwise from Build. Govern applies to every stage.

Each component is optional: use the ones your system needs at each stage. A workflow with no evaluation yet still deploys, and you can add evaluations later. Testing, observing, and improving rely on observability and evaluations, which are available to Enterprise-tier organizations only.

Before you design a system, read what happens at each stage of the loop, which components serve it, and how to choose between an agent, a workflow, and an app.

Pick the system closest to your project and follow its quickstart. Every Studio quickstart is also listed in Quickstarts on the home page.

Every component has its own quickstart, by stage of the loop:

• Set up: create your first API key, install the Mistral CLI and create a project.
• Build: build your first workflow, build your first app.
• Context: add a connector, create a library and ask it a question, create and query an index, build a RAG pipeline, create and use a prompt.
• Test and evaluate: run your first evaluation.
• Deploy: deploy a workflow to Mistral.
• Observe and improve: send your first trace, score your first span with online evaluations.
• Govern: moderate an input.
• Voice: transcribe an audio file, transcribe speech in realtime, generate speech, build a voice agent.

A coding agent, such as Vibe, Claude Code, or Codex, can build Studio systems for you. The Mistral CLI gives it commands to sign in, pick the workspace, and scaffold and run apps and workflows, and the Mistral MCP server exposes your workspace as tools.