Mistral Workflow Builder: Power-User Guide to Vibe and HTTP APIs

On November 5, 2026, Mistral added a workflow builder to Studio in public preview. In Vibe Code Web, you describe a process in plain language; the agent writes inspectable Python workflow code, tests it, and deploys it as a managed worker. The same release added HTTP connectors that turn an OpenAPI specification into tools agents and workflows can call.

Last updated: November 8, 2026

Table of contents

What changed on November 5, 2026

The Mistral release notes list the workflow builder as a Studio feature in public preview. You describe a process in natural language; Mistral Vibe writes a workflow, tests it, and deploys it, and you can inspect and edit the generated code. The same notes add HTTP API connectors, so a web API can become a connector without an MCP server, and they mark Studio apps in Vibe Code Web as generally available, with the workflow graph beside the conversation.

Mistral’s Studio announcement says an agent investigates and acts, a workflow coordinates execution and approvals, and an app gives employees a place to use those systems. The company says a product-operations workflow that routes ideas and feedback saves that team about 35 hours a week. The same day also shipped managed agents.

When to use a workflow

Mistral’s chooser says to pick by the shape of each part of a process. Use a workflow when you can write the steps as business rules ahead of time, an agent when the next step depends on what the system finds, and an app when people need a place to review results or approve an action. People can use agents and workflows in Vibe before you build an app. Observability and evaluations are available to Enterprise-tier organizations only. Always-on research agents such as Perplexity Teammates sit on the agent side of that split.

How to build a workflow in Vibe

You do not need a local toolchain for the first version. Build in Vibe is for domain experts who can describe a process, and for developers who want to skip scaffolding and then refine the generated Python.

  1. Open Vibe Code Web and describe the process in plain language, including inputs, systems, approvals, and failure paths.
  2. Watch the agent plan. The workflow graph appears beside the conversation, so you can correct a branch before anything is deployed.
  3. Ask the agent to adjust anything that is not right. The generated workflow stays editable at each step.
  4. Let the agent deploy the workflow to your workspace when you are happy with it.

Vibe Code Web sessions already include the Mistral MCP server tools that deploy, execute, and debug workflows. Other MCP clients connect to https://api.mistral.ai/mcp over streamable HTTP with Authorization: Bearer <MISTRAL_API_KEY>. See the Vibe CLI 3.0 guide if you add that server from the CLI. A workflow built in Vibe deploys as a managed deployment: Mistral builds and runs the worker, and the code lives in your Git repository.

Plans and managed deployment quotas

Managed deployments are in public preview. You point Mistral at a GitHub repository; Mistral Cloud builds a Docker image from the repository’s Dockerfile and runs the container. You do not manage servers or create an API key for the worker.

Plan Concurrent managed deployments
Free 0
Pro or Team, pay-as-you-go off 0
Pay-as-you-go enabled 3
Enterprise 10

The quota page says managed deployments require pay-as-you-go or an Enterprise plan. Creating a deployment outside those plans, or after you hit the cap, fails with a 403 error until you stop or delete a deployment. Mistral Cloud picks one instance size for every managed worker and does not expose per-deployment CPU or memory metrics. Pair a builder workflow with Mistral Large 4 when you need the current flagship model.

HTTP connectors from an OpenAPI spec

Not every API has an MCP server. An HTTP connector imports the API’s OpenAPI specification and creates one tool per supported operation, keeping the full operation surface visible for workflows, conversations, and agents.

  1. Collect the server URL (for example https://api.spotify.com/v1), the authentication method (none, an API key, or OAuth), and an OpenAPI file or public URL in JSON or YAML. Studio supports OpenAPI 3.0, 3.1, and 3.2, not Swagger 2.0.
  2. In Studio, go to Connectors, click + Add connector, choose Add a custom connector, select the HTTP protocol, and enter the server URL.
  3. Choose authentication. For OAuth, enter the authorization endpoint, token endpoint, and scopes, or paste the API’s OAuth server metadata. Wrong scopes fail the consent screen, and the import skips operations that need a missing scope.
  4. Click Create. With OAuth, Studio redirects you to the API’s consent screen.
  5. Upload the specification or paste a public URL, review convertible operations, then click Validate and add tools.
  6. Turn individual tools on or off, then use the connector as you would an MCP connector. Names accept up to 64 letters, digits, underscores, and dashes. Visibility can be private, workspace-shared, or organization-shared (organization admins only for org-wide connectors).

In workflow code, the connector integration (mistralai-workflows-plugins-mistralai, public preview) declares slots with connector("name"), injects a ToolCallClient, and attaches them with @uses_connectors(...) after @workflow.define. Regular workflows use the worker’s credentials; on-behalf-of workflows use the triggering user and need a hardened deployment. A missing OAuth credential pauses the run, emits an auth_url, and waits up to 10 minutes before ConnectorAuthTimeout.

Event triggers and webhooks

You can also start a workflow from an external event. Create a webhook from the Webhooks page or POST /v1/webhooks, deploy a workflow that accepts the incoming HTTP request as its input, then create a trigger with handler workflows and action type start_workflow. Use the workflow name as primitive_id, not the workflow ID. If more than one deployment is active, the trigger fails to start the workflow. Built-in providers include Slack, GitHub, Notion, Linear, Microsoft Graph (microsoft), Lucca, and Ashby. Optional JMESPath filters can restrict matching events. Triggers cannot start OBO workflows, and the run uses the identity of the user who created the trigger. Delivery is asynchronous and deduplicated, and transient failures are retried up to four times after 60, 180, and 600 seconds.

Frequently asked questions

What is the Mistral workflow builder?

It is a public-preview Studio feature announced on November 5, 2026. In Vibe Code Web, you describe a process in plain language; the agent writes, tests, and deploys a workflow, and you can inspect and edit the generated code at every step.

What plan do I need to deploy a Mistral workflow?

A workflow built in Vibe deploys as a managed deployment, which is in public preview. You need pay-as-you-go enabled or an Enterprise plan. Free, and Pro or Team without pay-as-you-go, have a cap of 0 and return a 403 error. Pay-as-you-go allows 3 concurrent managed deployments; Enterprise allows 10.

What are Mistral HTTP connectors?

An HTTP connector imports an API’s OpenAPI specification into Studio and creates one tool per supported operation, so agents and workflows can call the API without an MCP server. Studio supports OpenAPI 3.0, 3.1, and 3.2, not Swagger 2.0. Authentication can be none, an API key, or OAuth.

Can a trigger start an on-behalf-of workflow?

No. Triggers cannot start on-behalf-of (OBO) workflows. The workflow runs as the user who created the trigger, not as the person behind the provider event. Use the workflow name, not its ID, and keep only one intended deployment active.

How does a Mistral workflow differ from a managed agent?

In a workflow, your code decides every next step, and Studio keeps execution state and retries failed steps. In an agent, the model decides the next step from its instructions, context, and tools. Mistral describes agents as investigating and acting, and workflows as coordinating execution and approvals. They can share connectors and secrets.

Sources

Image: Mistral AI logo, public domain, via Wikimedia Commons. Placed on a white background and resized. Mistral AI is a trademark of Mistral AI. FreeAnswers is not affiliated with or endorsed by Mistral AI.

Leave a Comment