Mistral Studio
Mistral AI · Mistral Studio: Enterprise Platform for Managed Agents, Durable Workflows and Apps on Mistral Models
Mistral Studio is Mistral AI's developer platform for building, deploying, and governing agents, workflows, and internal apps on Mistral models. Its Managed Agents (public preview since November 5, 2026) are Python agents built with the Agents SDK that run for hours or days, resume after a crash, wait for human approvals, and execute code in sandboxes, either self-hosted with agents serve or Mistral-hosted with agents deploy. It suits enterprise engineering teams, especially in regulated industries, that want a Mistral-based agent runtime with tracing, evaluations, and private deployment options.
Best for
Enterprise engineering and platform teams that want to build Python agents and workflows on Mistral models with durable sessions, approvals, sandboxes, tracing, and evaluations, and that may need to keep tools, data, or the whole platform inside their own network
Not ideal for
Teams that want a model-agnostic framework, non-Python developers, anyone who needs a stable GA API today, buyers who need published per-agent or sandbox pricing, and non-technical users looking for a ready-made assistant
Who it's for
Enterprise developers, platform teams, and systems integrators building agents, workflows, and internal AI apps on Mistral models, particularly in regulated sectors such as banking and the public sector
Capabilities
- Managed Agents defined in Python with the Agents SDK (mistralai-agents, Apache-2.0): model, instructions, typed tools, connectors, MCP servers, skills, secrets, guardrails, hooks, and structured output in one agents.Agent(...) definition
- One codebase, three run modes: in your own process for development, self-hosted workers with agents serve, or Mistral-hosted with agents deploy, which builds a pushed GitHub commit from the project's Dockerfile
- Durable deployed sessions that can run for days, resume from the last finished model or tool call after a worker crash, and are stored by Mistral, so you host no database
- Approvals per tool call (permission_mode and permission_policy), answered by your code or left for a person, with the turn waiting even while your app is closed (one hour by default)
- Steer or interrupt a running turn, and connect several clients to one session for live views and approvals
- Subagents that run side by side on parts of a task, with one approval policy covering the whole tree and limits on concurrency and depth
- Sandboxes: a Mistral-hosted Linux container per session with Python, a shell, and a writable workspace, or local Docker and in-process sandboxes for development and CI, with outbound network policies
- Skills as SKILL.md bundles loaded on demand from your project or Studio's AI Registry, and hooks that run your code at fixed points of a turn
- Mistral connectors that act with the session user's own permissions, plus your own remote MCP servers with credentials read from environment variables, and HTTP API connectors (private preview)
- Workflows built on Temporal for multi-step processes with retries, scheduling, human signals, and crash recovery, in hybrid mode where your workers run the code, with Mistral-hosted managed deployments as an option
- Workflow builder in Vibe that writes, tests, and deploys a workflow from a plain-language description, with editable code and a live workflow graph
- Apps: mistral apps commands scaffold, run, and deploy internal apps with hosting, authentication, and access controls, available at their own link or inside Mistral Vibe
- Tracing in Studio on by default, with one span per model and tool call and default redaction of common secrets and personal data before spans leave the worker
- Offline evaluations from the mistralai-evaluations SDK, online evaluations on live traffic, and datasets built from CSV, JSONL, or production traces
- Webhooks and triggers that start an agent session or workflow execution when something happens in your tools, plus a Secrets Manager with workspace RBAC and versioned secrets
- Mistral MCP server so Mistral Vibe or another coding agent can build against Studio and deploy with one command
Limitations
- Managed Agents, the Agents SDK, Apps CLI, workflow builder, triggers, managed search index, managed-agent tracing, and evaluations launched as public previews on November 5, 2026. For Managed Agents and the SDK, Mistral warns that the API, behavior, and limits can change during the preview and may require code changes
- Mistral's Workflows documentation still shows a Public Preview notice, although launch coverage described Workflows as generally available, so check the status for your account
- The Agents SDK is Python-only (Python 3.12 or later), and Mistral-hosted deployment builds from a GitHub repository through the GitHub connector
- Mistral-hosted managed deployments require pay-as-you-go or an Enterprise plan. Mistral's quota page lists 0 concurrent managed deployments on Free and on Pro or Team without pay-as-you-go, 3 with pay-as-you-go, and 10 on Enterprise
- Mistral's public pricing pages list per-token model prices but no itemized prices for Mistral-hosted sandboxes or managed deployments
- Self-hosted workers (agents serve) still rely on Mistral's control plane to store sessions and on Mistral's API for model calls, so they are not an air-gapped option
- Running the whole platform in a private cloud or on premises is for enterprise customers, and launch coverage reports private-cloud and on-premises availability for the new release only from the end of November 2026
- Deployed sessions end after 24 hours or after 4 hours idle by default (both can be raised), and a tool call interrupted by a worker crash may run twice, so tools with side effects must be idempotent
- Only the default Mistral sandbox region (nl-north-1) can restrict outbound traffic. Sandboxes in other regions must allow all network access
- Built around Mistral's models: Managed Agents call models through Mistral's API (or your own gateway), and the documented examples use Mistral Large 4 and Mistral Medium
- Mistral says the older Agents and Conversations API (Agents API v1) is about to be deprecated, so existing integrations should plan a move to the Agents SDK
Use cases
- Deploying an incident-investigation agent that follows the team's runbook when an alert fires, asks before sensitive actions, and posts findings in Slack
- Running a payment-review or KYC agent that gathers account history and sanctions data on its own, then waits for an analyst's approval before opening a case
- Serving an agent on your own servers with agents serve so its tools reach internal systems while sessions stay durable
- Generating a feedback-triage workflow from a plain-language description and routing bug reports and feature requests to the right team
- Shipping an internal staffing or document-review app with hosting, authentication, and access controls handled by Studio
- Testing a prompt, model, or tool change against your own dataset before release, then monitoring live traffic with online evaluations
Our take
Studio is Mistral's answer to managed agent runtimes from the big clouds and model labs, and its strongest argument is control. The same Python agent can move from a laptop to your own workers to Mistral's cloud, and Mistral says the models and platform can run inside your perimeter, which matters to the European banks and industrial firms Mistral is courting. The design is thoughtful: durable sessions, approvals that outlive the app, and tracing on by default cover what most teams end up building by hand. Today, though, it is a public preview with a Python-only SDK, and the self-hosted worker mode still depends on Mistral's control plane and API. Hosted sandbox and deployment costs are also not itemized publicly. Teams already committed to Mistral models should prototype now. Teams that need GA stability, model choice, or fully on-premises operation should wait for the private-cloud and on-premises release, which launch coverage reports for the end of November, and confirm pricing with Mistral.
Who should use it
Python teams standardizing on Mistral models who want durable, approval-gated agents with tracing and evaluations, platform teams that need agent tools to run inside their own network, and enterprises in regulated sectors that plan to run Mistral's models and platform in their own VPC or on premises.
Who should skip it
Teams that want a multi-model, open-source framework they fully control, developers working outside Python, organizations that cannot build on preview APIs, and buyers who need a ready-made assistant rather than a platform for building one.
Strengths
- The same agent code runs locally, on your own workers, or Mistral-hosted
- Durable sessions survive crashes and can wait on human approvals without a running app
- Tracing and evaluations are built in rather than bolted on
- Self-hosted workers keep tools and data access inside your network
- Mistral says the models and the platform can run in your own VPC or on premises for enterprises
Weaknesses
- Most new capabilities are public previews whose APIs can change
- Python-only SDK, built around Mistral's models
- Hosted deployments need pay-as-you-go or Enterprise, and sandbox and hosting prices are not itemized publicly
- Full private-cloud and on-premises deployment is enterprise-only and, for this release, reported for end of November 2026
Mistral Studio pricing
Free
$0
- Access to Mistral Studio on Mistral's Free plan
- Limited usage shared across Studio, the API, and Vibe
- No Mistral-hosted managed deployments (quota of 0)
Pro / Team
$14.99/mo (Pro) or $24.99/user/mo (Team)
- Monthly included usage shared across Studio, the API, and Vibe Code
- No Mistral-hosted managed deployments unless pay-as-you-go is enabled
Pay-as-you-go credits
Prepaid, usage-based
- Prepaid credits on top of a plan, used at API rates after included usage runs out
- Mistral Large 4: $0.70 input, $0.07 cached input, $2.10 output per million tokens
- Mistral Small 4: $0.15 input, $0.60 output per million tokens
- Up to 3 concurrent managed deployments
Enterprise
Custom
- Private deployments with custom models, agents, and workflows
- Up to 10 concurrent managed deployments
- SAML SSO and audit logs
- Custom SLAs and dedicated support
Free tier limits: The Free plan includes Studio access, but Mistral's quota page lists 0 concurrent managed deployments on Free, and on Pro or Team plans unless pay-as-you-go is enabled.
Note: Mistral does not list a separate per-agent or platform fee for Studio on its public pricing pages. One Mistral plan covers Vibe, Studio, and the API, and pay-as-you-go is a prepaid credit balance added to a plan rather than a separate plan. Model usage is billed per token, batch processing is discounted 50%, and cached input tokens cost less. Charges for Mistral-hosted sandboxes and managed deployment compute are not itemized on the pricing pages reviewed, and Enterprise private, VPC, and on-premises deployments are priced by contract. Rates checked against Mistral's API pricing documentation on November 6, 2026.
Technical specs
API pricing
Per-token model pricing through the Mistral API (for example Mistral Large 4 at $0.70 per million input tokens and $2.10 per million output tokens). Hosted deployments require pay-as-you-go or Enterprise
Available models
Where Mistral Studio excels
Incident investigation with a person in the loop
Mistral describes an internal agent that starts investigating when an alert fires, follows the runbook, asks before sensitive actions, and posts findings in Slack. Durable sessions mean the approval can wait without a running process.
Regulated reviews such as payment monitoring or KYC
An agent can gather account history, KYC profiles, and sanctions data on its own, then pause for an analyst's decision, with every step stored in the session for later audit.
Agents whose tools must stay inside the corporate network
With agents serve, the worker and its tools run on your infrastructure next to internal systems, while Mistral stores the session so a crash does not lose finished work.
Mistral Studio vs. competitors
Mistral Studio vs. Google ADK
Google ADK is an open-source agent framework with SDKs for Python, TypeScript, Go, Java, and Kotlin, optimized for Gemini but supporting other models, which you deploy to Google Cloud's agent runtime or your own containers. Mistral Studio's Agents SDK is Python-only and built around Mistral models, but it comes with a Mistral-run control plane that stores sessions, resumes them after crashes, holds approvals open while your app is closed, and traces every turn by default. ADK fits multi-language teams that want framework-level control and model choice, and Studio fits teams that want the runtime managed and plan to stay on Mistral.
Mistral Studio vs. Microsoft Agent Framework
Microsoft Agent Framework is an MIT-licensed SDK for Python and .NET, the successor to AutoGen and Semantic Kernel, with graph workflows, checkpointing, and wide model support, and it can deploy to Microsoft Foundry Hosted Agents. Mistral Studio pairs a Python SDK with its own hosted runtime, sandboxes, workflows, apps, and evaluations, all built around Mistral models and Mistral's private-deployment options. Microsoft Agent Framework suits .NET and Azure shops that want model flexibility, and Studio suits teams that want a single Mistral platform from model to deployed agent.
Mistral Studio vs. DigitalOcean Managed Agents
DigitalOcean Managed Agents runs agents you already use, such as Claude Code, Codex, OpenCode, or LangGraph and CrewAI apps, in Firecracker microVMs with pause, resume, and fork, a 16,000-tool MCP gateway, and published per-second compute rates, but only on DigitalOcean's cloud. Mistral Studio runs agents you write with its own Python SDK, adds workflows, apps, and evaluations, and offers a self-hosted worker mode plus enterprise private deployment. DigitalOcean fits developers hosting an existing harness, and Studio fits enterprises building their own agents on Mistral models.
Mistral Studio vs. TrueForge
TrueForge is an MIT-licensed, model-neutral agent harness that you self-host, with sessions, approvals, subagents, MCP tools, and Daytona sandboxes, exposed through a chat UI, HTTP API, and TypeScript SDK. Mistral Studio covers similar runtime concerns (durable sessions, approvals, sandboxes, subagents) as a Mistral service with a Python SDK, adding workflows, apps, tracing, and evaluations, but it is built around Mistral's models and depends on Mistral's control plane. TrueForge fits teams that want to own the whole runtime and keep model choice open, and Studio fits teams that prefer a vendor-run platform on Mistral models.
Frequently asked questions
What is Mistral Studio?
Mistral Studio is Mistral AI's developer platform. Besides the console for API keys, the Playground, and usage tracking, it covers building and deploying managed agents, durable workflows, and internal apps on Mistral models, with connectors, secrets, tracing, evaluations, and governance controls.
What are Mistral Managed Agents?
Managed Agents are agents you write in Python with the Agents SDK (mistralai-agents) and deploy with Studio. Once deployed, a session can run for hours or days, resume from its last finished step after a crash, and pause for a person's approval. Mistral stores the session and traces every model and tool call. Managed Agents and the SDK have been in public preview since November 5, 2026.
How do you deploy an agent in Mistral Studio?
During development the agent runs in your own process. For production you either run agents serve, which starts a worker on your own machine or cluster while Mistral stores sessions, or agents deploy, which has Mistral build a pushed GitHub commit from your Dockerfile and run it for you. The agent code and session calls are the same in all three modes.
How much does Mistral Studio cost?
Mistral's Free plan includes access to Studio, and model usage through the API is billed per million tokens. Mistral Large 4, for example, is $0.70 per million input tokens and $2.10 per million output tokens. Mistral-hosted managed deployments need pay-as-you-go credits enabled on the organization (up to 3 running at once) or an Enterprise plan (up to 10), and Enterprise pricing for private deployments is custom. Mistral's public pricing pages do not itemize charges for hosted sandboxes or managed deployments.
Can Mistral Studio be self-hosted?
Partly today. With agents serve, your agent's worker and tools run on your own infrastructure, but sessions are stored by Mistral and model calls go to Mistral's API. Workflow workers likewise run in your environment while Mistral hosts the orchestrator. Mistral says enterprise customers can run both the models and the Studio platform in their own VPC or on premises, and launch coverage reports that private-cloud and on-premises availability for the November 2026 release is due at the end of November.
Is the Mistral Agents SDK open source?
The mistralai-agents package on PyPI is published under the Apache-2.0 license and requires Python 3.12 or later. Studio itself, including the hosted control plane, sessions storage, and tracing, is a Mistral service.
Mistral Studio vs Mistral Vibe: what is the difference?
Mistral Vibe is Mistral's ready-made assistant and coding agent for end users, on the web, mobile, the terminal, and IDEs. Studio is the developer platform for building your own agents, workflows, and apps. They connect: Vibe can generate Studio workflows, a deployed Studio agent can be tried in Vibe Code Web, and Studio apps can be made available inside Vibe.
Mistral Studio vs Google ADK: what is the difference?
Google ADK is an open-source framework with SDKs in five languages, optimized for Gemini but able to run other models, that you deploy to Google Cloud or your own infrastructure. Mistral Studio pairs a Python-only SDK with a Mistral-run control plane that stores sessions, resumes them after crashes, and traces every step, and it is built around Mistral's own models.
Integrations & fit
Alternatives to consider
Developers building production agents in Python, TypeScript, Go, Java, or Kotlin who want an open-source, code-first framework with first-class Gemini support and a direct path to Google Cloud
FreeMicrosoft
Python and .NET multi-agent orchestration, especially with Azure and Microsoft Foundry
Free
DigitalOcean
Developers and small teams who already use Claude Code, Codex, OpenCode, OpenHands, Hermes, LangGraph, or CrewAI and want those agents running durably in the cloud, with pause, resume, and fork, governed tool access, and usage-based billing in one place
Paid
TrueFoundry
Platform and application teams that want a model-neutral, self-hosted runtime for reusable production agents
FreeAbout Mistral Studio
Studio is Mistral's developer console (API keys, Playground, usage tracking) and, increasingly, its platform for running AI inside a business. On November 5, 2026 Mistral expanded it with Managed Agents, Apps, and new workflow, observability, and evaluation features. The pieces split one business process three ways. A managed agent investigates and acts, a workflow fixes the order of steps, retries, and approvals, and an app gives employees a front end. For Managed Agents, you write the agent in Python with the Apache-2.0 Agents SDK (mistralai-agents). You set its model, instructions, typed tools, connectors and MCP servers, skills, secrets, guardrails, hooks, subagents, sandbox, and which tool calls need a person's approval. The same code runs three ways: in your own process while you build, as a self-hosted worker (agents serve) that keeps your tools inside your network while Mistral stores sessions, or Mistral-hosted (agents deploy), which builds a pushed GitHub commit from its Dockerfile. Deployed sessions are durable. A turn resumes from its last finished model or tool call after a crash, an approval can wait while your app is closed, and each turn shows up as a trace in Studio with no setup. Workflows, built on Temporal, cover processes with explicit steps. They run in hybrid mode, where Mistral hosts the orchestrator and your workers run the code, and they can also be generated from a plain-language description in Vibe. Around both sit managed search indexes, webhooks and triggers, offline and online evaluations, datasets built from traces, and a Secrets Manager. Two things set Studio apart. It is built around Mistral's own models, starting with Mistral Large 4 ("Le Chonk"), and Mistral says enterprises can run both the models and the platform in their own VPC or on premises. Launch coverage reports that private-cloud and on-premises availability for this release is due at the end of November 2026. The tradeoffs follow from how new it is. Most of the new pieces are public previews whose APIs can change, the SDK is Python-only, Mistral-hosted deployments need pay-as-you-go or an Enterprise plan, and sandbox and hosting charges are not itemized on Mistral's public pricing pages.
Updates from Mistral Studio
Mistral expanded Studio with Managed Agents and the Agents SDK (public preview), which run Python agents in your process, on your own workers, or Mistral-hosted, with durable sessions, approvals, sandboxes, skills, subagents, and hooks. The same release added the Apps CLI, a workflow builder in Vibe, webhooks and triggers, a managed search index, tracing for managed agents, offline and online evaluations, and managed datasets, mostly in public preview, plus HTTP API connectors in private preview.
Studio added a Secrets Manager for API keys, tokens, and credentials, scoped to a workspace with role-based access control, versioned for rotation, audited, and injected into managed deployments at runtime.
Managed deployments let teams run workflow workers on Mistral Cloud from a GitHub repository instead of operating their own workers. During the public preview they require pay-as-you-go or Enterprise and are subject to a concurrency quota.
Are you the founder? Claim this listing →