Mistral Vibe

Mistral Vibe is a unified AI assistant and agent from Mistral AI for professional work, software development, and conversational tasks. The current product spans web and mobile apps, a coding agent for the terminal and Visual Studio Code, and managed cloud coding sessions. Mistral organizes it into Work, Code, and Chat modes.^[1]

The name originally referred only to an open-source command-line coding agent. Mistral released Vibe CLI with the Devstral 2 model family on December 9, 2025, then expanded the name in May 2026 when it renamed Le Chat to Vibe and combined work automation, coding, and chat under one product and subscription.^[2][4] The command-line client remains open source, but the hosted Vibe service and its cloud infrastructure are commercial products.

History

Mistral introduced Vibe CLI as a terminal-native coding agent built for Devstral. The first release could inspect a project, search and edit files, run commands, use version-control tools, preserve session history, and operate programmatically. It was also designed to work through editor clients rather than only through its own terminal interface.^[2] TechCrunch described the launch as Mistral's entry into the command-line coding-agent market and noted that the smaller Devstral model could be run on consumer hardware, while the 123-billion-parameter Devstral 2 required substantially more infrastructure.^[16]

Vibe 2.0 followed on January 27, 2026. It added custom subagents, clarification prompts, reusable Skills exposed through slash commands, several approval modes, and automatic client updates. Mistral included it with its Pro and Team subscriptions, with pay-as-you-go overflow or bring-your-own-key access.^[3]

In May 2026, Mistral added remote coding agents and a Work mode for multi-step tasks across documents and connected applications. Its May 28 announcement stated that Le Chat was now Vibe and that existing conversations, settings, and plans would carry over.^[4] Current documentation describes the result as one agent across professional work and coding, with the older turn-based assistant retained as Chat mode.^[1]

Work, code, and chat

The three modes share an account but serve different workflows:^[5]

Scroll sideways for more →

Work can gather context from uploaded files, the web, libraries, and connectors. Mistral advertises integrations with more than 100 tools, including email, calendars, Slack, GitHub, Jira, and document stores. The agent displays progress and requests approval before sensitive actions such as sending a message, deleting data, or changing a connected service.^[5][18] Reusable instructions can be packaged as Agent Skills, while scheduled tasks can run prompts or workflows at set times.

Code can work against a local checkout or a remote repository. Local sessions run through the CLI or VS Code extension. Vibe Code Web creates a managed Linux sandbox, clones an authorized repository, runs the agent, and leaves a branch, commit, or pull request for review. A terminal session can also send a task to the cloud or transfer its context to a web session.^[5][8]

Chat preserves the earlier conversational interface while Mistral moves or redesigns features for Work. The two modes are not exact substitutes. For example, the legacy Chat Code Interpreter runs Python, while Work uses a TypeScript environment for many data-processing and charting tasks.^[5]

CLI and extension system

The official mistralai/mistral-vibe repository is licensed under the Apache License 2.0. Manual installation requires Python 3.12 or later, with packages distributed through uv and pip; Mistral also supplies a shell installer for macOS and Linux. The project works on Windows, although its repository says Unix environments are the official support target.^[6][7]

In an interactive terminal session, a user can attach files with @, open slash commands with /, or prefix a command with ! to execute it directly in the shell. Built-in tools cover file reading and editing, code search, shell execution, todo tracking, clarification questions, and task delegation to subagents. Agent profiles range from read-only planning to auto-approval. Projects can provide configuration, custom agents, prompts, hooks, and Skills in a .vibe directory.^[6]

The CLI supports servers that implement the Model Context Protocol, including local standard-input servers and hosted HTTP services. It can also run through the Agent Client Protocol in compatible editors such as Zed. Vibe's VS Code extension contains its own agent, but it shares configuration and sessions with the CLI.^[6][7]

Programmatic mode supports turn, token, cost, and tool limits plus text, JSON, or streaming output. There is an important documentation conflict around approvals. Mistral's web documentation says non-interactive runs default to auto-approval, but the project's changelog records that version 2.12.0 stopped doing so on May 27, 2026, and the current README says programmatic sessions use the configured default agent unless --auto-approve or --yolo is passed.^[12][13] Users should select an agent explicitly and verify the behavior of the installed release instead of assuming either default.

Models and deployment

Vibe is a client and agent system rather than a single model. The original CLI was built around Devstral 2 and Devstral Small 2, and Vibe 2.0 continued to use that family.^[2][3] Current Vibe Code Web sessions use Mistral Medium 3.5. A local CLI session can use a selected Mistral model, a fully local model, or a compatible third-party API. If a session using a non-Mistral model is transferred to the cloud, Mistral's limits page says the web session defaults to Medium 3.5.^[7][8]

This distinction affects privacy, hardware, and cost. A fully local model can keep inference on the user's system, but its memory and accelerator requirements come from that model, not from Vibe itself. A hosted model sends the selected context and tool results to the configured provider under that provider's terms. Vibe Code Web instead runs in Mistral's infrastructure.

Enterprise customers can deploy Vibe on premises, in a private cloud, or on Mistral Cloud, according to Mistral's product page. Mistral also offers model customization for enterprise deployments.^[18] These options do not make every component open source, and the company does not publish one universal architecture, hardware requirement, or price for enterprise installations.

Mistral Code relationship

Mistral Vibe is not simply a rename of Mistral Code, an enterprise coding assistant launched in June 2025. Mistral Code was based on Continue and combined several specialized models for completion, retrieval, agentic coding, and chat. It also emphasized administrative controls and private enterprise deployment.^[9]

Mistral's migration guide treats the products as separate extensions and separate plans. It says they can coexist, but settings and saved sessions do not transfer automatically. The guide contrasts Mistral Code's multi-model arrangement with Vibe's single-agent session design and directs enterprise customers to contact their account team when adding Vibe.^[10] Vibe is therefore the company's newer unified coding-agent path, while Mistral Code remains a distinct legacy enterprise product.

Access and pricing

As of July 2026, Mistral lists a free Vibe tier with limited messages, searches, and coding sessions. Pro costs USD 14.99 per month, and Team costs USD 24.99 per user per month, before tax and subject to fair-use limits. An eligible student plan costs USD 5.99 for up to 12 months, while Enterprise uses custom pricing.^[11]

Subscription access and API billing are separate. The CLI can use included plan access, pay-as-you-go capacity where available, a separately managed API key, a third-party provider, or a local model. Vibe Code Web currently allows two sessions per day for free users where access is enabled and 100 per day for paid users. Concurrent-session limits depend on plan.^[7][8]

Security and data handling

Local Vibe uses layered controls. The default agent asks before tools run, the plan agent is read-only, accept-edits can approve file changes while still asking about commands, and auto-approve permits all tool activity. Per-tool rules can further limit access. Vibe also asks before activity outside the working directory and requires trust before loading repository-level .vibe configuration. Mistral advises reviewing such configuration because it can define commands, external tools, and credentials.^[12]

The open-source client collects usage telemetry and crash reports unless telemetry is disabled in its configuration.^[6] That control applies to the client and should not be confused with the data policy for hosted Vibe. Mistral says free-plan Vibe conversations can be used to improve its models unless the user opts out. Pro, Team, and Enterprise conversations are not used for training by default, and API data is not used for training. Connector data is fetched on demand, is not stored permanently, and is not used for training. Uploaded library documents are stored but are not used for training.^[14]

Each Code Web session runs in a single-tenant sandbox that is deleted when the session ends. It has outbound internet access, no public inbound access, and credentials for authorized repositories. Files and build artifacts disappear unless committed, but prompts, agent messages, command output, and change history are retained according to the user's plan.^[8][15]

Mistral explicitly identifies prompt injection as a risk when the agent reads repository files, command output, web pages, issues, or pull-request comments. The activity stream makes actions visible, and normal branch protections, required reviews, status checks, and CODEOWNERS rules still apply. These controls do not replace review of the resulting code.^[15]

Evaluation and limitations

At launch, Mistral reported that Devstral 2 scored 72.2% on SWE-bench Verified and that Devstral Small 2 scored 68.0%. It also commissioned human-preference comparisons using tasks scaffolded through Cline.^[2] These are evaluations of the models and a particular agent scaffold, not a benchmark of the current Vibe product across Work, Code, and Chat. Mistral has not published a standardized product-level success rate for Vibe.

A July 2026 TechRadar review based on several weeks of Pro-plan use found Work effective for multi-step research and document drafting. Its Code test scaffolded an API integration, ran tests, and prepared a draft pull request, but required two interventions. The reviewer also criticized unclear usage limits and less native productivity-suite integration than some competing platforms.^[17] This is an independent hands-on assessment, not a controlled evaluation.

Vibe Code Web sessions have a 24-hour maximum duration and a three-hour timeout while waiting for input. Once a sandbox is removed, the session cannot be resumed, although pushed branches and pull requests remain. The standard environment cannot reach local files or services, is not intended for model training or heavy pipelines, and does not yet offer user-defined internet allowlists or custom network controls. Mistral recommends test or least-privileged credentials and warns against placing production secrets in the sandbox.^[8][15]

More generally, approval prompts and isolated environments limit the blast radius of errors but do not guarantee correct or secure output. Generated code can contain defects, shell commands can be destructive, connected tools can change external data, and untrusted context can redirect an agent. Work outputs and code changes still require factual checks, testing, security review, and human approval appropriate to their consequences.

References

1. Mistral AI. Vibe overview. Mistral Docs. ↩
2. Mistral AI. Introducing: Devstral 2 and Mistral Vibe CLI. December 9, 2025. ↩
3. Mistral AI. Terminally online Mistral Vibe. January 27, 2026. ↩
4. Mistral AI. Vibe gets to work. May 28, 2026. ↩
5. Mistral AI. Choose Chat, Work, or Code. Mistral Docs. ↩
6. Mistral AI. mistralai/mistral-vibe. GitHub repository. ↩
7. Mistral AI. Install the Vibe CLI and send your first prompt. Mistral Docs. ↩
8. Mistral AI. Limits and lifecycle. Mistral Docs. ↩
9. Mistral AI. Introducing Mistral Code. June 4, 2025. ↩
10. Mistral AI. Migration from Mistral Code Enterprise. Mistral Docs. ↩
11. Mistral AI. Pricing. ↩
12. Mistral AI. Safety, approvals, and permissions. Mistral Docs. ↩
13. Mistral AI. Mistral Vibe changelog. GitHub. ↩
14. Mistral AI. Privacy and data controls. Mistral Docs. ↩
15. Mistral AI. Sandbox environment and Security. Mistral Docs. ↩
16. Anna Heim. Mistral AI surfs vibe-coding tailwinds with new coding models. TechCrunch, December 9, 2025. ↩
17. TechRadar Pro. Mistral Vibe review. July 10, 2026. ↩
18. Mistral AI. Mistral Vibe. Product page. ↩