Mistral Vibe (Le Chat)

Mistral Vibe is Mistral AI’s assistant and agent for work and coding, incorporating the experience formerly called Le Chat. It combines conversation, research and deliverable creation with connected tools when those tools are enabled.

In brief

AI assistants

Official website

Contents
  1. What Mistral Vibe is
  2. Le Chat's origins and early expansion
  3. News, mobile applications and memory in 2025
  4. From coding agent to shared brand
  5. How a request becomes work
  6. Search, sources and Deep Research
  7. Documents, audio and Canvas
  8. Spreadsheets and indicators
  9. Mini apps and visualisation
  10. Images and voice
  11. Memory, projects and stable context
  12. Skills and recurring tasks
  13. Connections and external actions
  14. Coding, licensing and change control
  15. Plans, deployment and cost evaluation
  16. Personal data and training
  17. Documented security incident and updates
  18. Reliability and regulatory response
  19. Relationship to occupational safety and health
  20. Hypothetical case: weekly prevention agreements
  21. When to choose it and complementary tools
  22. Frequently asked questions
  23. Related profiles
  24. On the Sabentis blog
  25. Sources and documentation

Back to Tools

What Mistral Vibe is

Mistral Vibe is Mistral AI’s assistant for conversations, document work and software development. Its Work experience combines quick questions and multi-step tasks; Code focuses on repositories and programming. It is available on the web and mobile and, for development, in an editor and terminal.

Mistral AI began in 2023 and was founded by Arthur Mensch, Guillaume Lample and Timothée Lacroix. The company, its models and this application are different things: subscribing to an assistant, calling an API and deploying a model are separate decisions, even when they share a provider.

Reference documentation: Vibe overview · About Mistral.

Le Chat's origins and early expansion

Le Chat launched in beta on 26 February 2024 as a conversational interface to Mistral models. Its initial announcement warned that it could not consult the internet. Descriptions of that launch therefore do not represent its current operation.

On 18 November 2024, Mistral announced web search with citations, Canvas for editing outputs, document and visual analysis using Pixtral Large, and image generation with Black Forest Labs’ Flux Pro. Le Chat, named after the French word for cat, began extending beyond text answers.

Reference documentation: Le Chat, lanzamiento · Mistral has entered the chat.

News, mobile applications and memory in 2025

In January 2025, the AFP partnership was announced, bringing news information and access to its text archive dating from 1983; photographs, videos and infographics were excluded. Having a news source available does not mean every answer uses it.

The iOS and Android applications, Pro and Team tiers and Code Interpreter arrived on 6 February. Free’s promotion in France then offered mobile subscribers a year of Le Chat Pro. This describes historical distribution, rather than a guaranteed current offer.

Flash Answers highlighted generation speed: Cerebras reported more than 1,100 tokens per second for text queries using Mistral Large 2. That inference measurement is different from the total time needed to research a topic, consult tools or complete a report.

The May Le Chat Enterprise offering added enterprise search, libraries, connectors and hybrid deployments. July brought Deep Research, voice with Voxtral, reasoning with Magistral, Projects and image editing. September expanded MCP connectors and introduced Memories, retaining context between conversations under user controls.

Reference documentation: AFP: acuerdo y archivo · The all new le Chat · Cerebras brings instant inference · Introducing Le Chat Enterprise · Le Chat dives deep · Le Chat: MCP and Memories · Free y Le Chat Pro.

From coding agent to shared brand

The Vibe name already identified the coding agent. Vibe 2.0, announced on 27 January 2026 with Devstral 2, added configurable subagents, clarification and skills. On 28 May, Mistral extended Vibe to the whole assistant, retaining Le Chat conversations, settings and plans while bringing work and code together.

September 2026 began the gradual merger of Chat and Work, with Fast and Think and an administrator-managed Enterprise migration period. When interpreting a guide, check which interface it describes: an old screenshot is not a universal description of the current product.

Reference documentation: Terminally online Mistral Vibe · Vibe gets to work · Release notes.

How a request becomes work

Work receives a request, gathers context from files or enabled tools and performs steps towards an output. It can display intermediate tasks, tool calls and deliverables. The organisation determines which capabilities are available; a missing connection may require authentication.

For a prevention team, the useful input is not “improve safety” but a bounded assignment: compare two authorised versions of an instruction, identify differences and prepare questions for its owner. The output should make it possible to check what was read, what changed and what remains unresolved. Breaking down the request this way makes omissions easier to detect before they become decisions.

Reference documentation: Get started with Work.

Search, sources and Deep Research

Web search finds current information; Open URL reads a specific page and Deep Research organises an investigation with sources. Open URL does not automatically crawl a whole website or access private or paywalled pages. Dynamic pages may be incomplete, and search can retrieve outdated information.

For occupational safety and health, Sabentis proposes using it to prepare a source map recording the organisation, scope, date and question each source addresses. A link beside a claim is insufficient: open it and check whether it supports that precise claim. For a regulatory question, a news story can reveal a change, but the applicable version must be found in the relevant official publication.

Reference documentation: Search the web · Reuse work with Skills.

Documents, audio and Canvas

Work accepts documents, images and tabular data; it can also transcribe a recording and answer from it. Canvas supports editing text, data, code and presentations, preserving versions and exporting results. These are production capabilities, rather than certification that extraction is faithful.

A scanned table with repeated headers can create incorrect associations. When comparing meeting minutes, request a file and section reference for each agreement. If several people speak in a recording, an uncertain attribution should not become the definitive owner of an action. Keeping the original allows the summary to be corrected without reconstructing the entire meeting.

Reference documentation: Files and Canvas.

Spreadsheets and indicators

On paid plans, Vibe Work can create workbooks with multiple sheets and formulas, edit them in Canvas and download them as XLSX. Opening a workbook in Google Sheets creates an independent version: changes made there do not automatically return to Canvas.

A proposed use is a draft action tracker with an identifier, date and status. Before using it, check interpreted dates, blank cells, duplicates and formulas against known examples. Counting closed actions does not demonstrate that a risk has decreased: the indicator must describe exactly the activity being measured. Agree which file will be the reference version to avoid maintaining two incompatible trackers.

Reference documentation: Spreadsheets.

Mini apps and visualisation

Canvas can generate React mini apps, expose their code and download their files. Product release notes describe them as snapshots, without continuous data refresh or persistent actions. Public links provide access to the shared application.

A revision quiz or filter demonstration can support training. However, a mock-up showing open tasks should not be presented as an operational register if it neither saves changes nor retrieves current information. Before sharing it, test mobile readability, contrast, keyboard use and what happens when a value is missing.

Reference documentation: Mini apps · Release notes.

Images and voice

Image generation and editing use Black Forest Labs models; small details or text can change. Work dictation converts speech to text and allows review before sending when Autosend is disabled. The documented language list includes Spanish and English; it does not establish an equivalent guarantee for Catalan here.

For a prevention campaign, a generated illustration can help explore composition or tone. If it shows equipment, protective devices or a specific operation, check every element against the real materials and approved instruction. A convincing image can teach an incorrect practice. Voice can help draft notes, but it does not eliminate mistakes in names, figures or negations.

Reference documentation: Generate images · Voice mode.

Memory, projects and stable context

Knowledge Base retains context in entries that can be reviewed and edited. Sensitive data has specific controls; disabling its storage does not erase previous entries. Memory does not authorise connector actions or replace their permissions.

For OSH, retaining formatting preferences, terminology or report structure is more reasonable than retaining workers’ personal information. A procedure change requires reviewing saved context: a template that was correct last month can produce outdated documents today. Personal context should also be distinguished from information intended for the whole team.

Reference documentation: Knowledge Base.

Skills and recurring tasks

Skills package reusable instructions and supporting files for personal or shared use. They can define an extraction method or template; they are not newly trained models. Scheduled tasks, in public preview, run requests at future times or on recurring schedules, with limits depending on the plan.

Sabentis proposes creating a weekly-summary skill only after reviewing several examples. It should specify how to handle undated agreements, contradictions and missing fields. Scheduling should prepare a draft for review. A recurring time should not, by itself, turn a recommendation into a sent communication or a summary into a register update.

Reference documentation: Reuse work with Skills · Schedule tasks.

Connections and external actions

Connectors include services such as Outlook, SharePoint, Slack, Notion and GitHub, with authentication and administrative controls. Work distinguishes queries from actions that modify systems. Always allow permissions can remove further confirmations; stopping execution does not necessarily reverse an action already received by an external service.

Scope should be decided by function: reading minutes, drafting a summary, changing a task and sending an email are different operations. Reading and drafting are sufficient for an initial prevention-workflow trial. Recipients, attachments and final modifications require review appropriate to the organisation’s process, even if the agent completed its earlier steps correctly.

Reference documentation: Connect tools with Connectors · Safety and approvals.

Coding, licensing and change control

Vibe Code works with files and commands through the CLI, VS Code or a GitHub repository in a remote environment. Documentation states that file edits are applied without asking by default; the ask agent enables review of these too. The public CLI repository carries Apache 2.0 licensing, which does not automatically turn the whole platform and its models into one open product.

It can help prepare a data transformation or explain a script, but an error-free run does not validate the result. In a training file, for example, removing duplicate rows by name could delete legitimate attendance records. The developer must define record identity and test exceptions before integrating the change.

Reference documentation: Vibe Code · Mistral Vibe repository.

Plans, deployment and cost evaluation

The offering reviewed distinguishes Free, Pro, Team and Enterprise. Team adds shared-workspace management; Enterprise provides for customised deployment and support arrangements. Check limits and features for the actual account and service, without treating historical Le Chat prices as the current tariff.

In May 2026, The Decoder observed that expressing some quotas as multiples of the free tier made absolute capacity difficult to establish. That historical observation encourages evaluating actual use: documents to process, task duration, storage, administration and the cost of reviewing each result. A generous quota does not compensate for a workflow that creates more corrections than useful work.

Reference documentation: Mistral Vibe product · The Decoder: cambio a Vibe.

Personal data and training

Mistral’s help documentation allows opting inputs and outputs out of training, including supplied documents. Enterprise is opted out by default; other environments have personal or administrative settings. Vibe and API controls are separate. According to documentation, Temporary Chat avoids history and training, with possible legally required retention.

Before handling OSH information, the organisation should establish what data may enter, where it is processed and who can access it. A European provider or a temporary conversation does not automatically settle those questions. A pilot can use fictional minutes or aggregated data, without medical records, identifiers or details that make someone recognisable.

Reference documentation: Training opt-out · Vibe overview.

Documented security incident and updates

Advisory MAI-2026-003 documents six vulnerabilities in CLI permission checks. Certain commands could bypass expected confirmation and access or modify files. Version 2.25.4, released on 12 September 2026, incorporated fixes. Running an affected version indicated potential exposure, rather than proving a compromise.

This case shows why evaluation must include the installed version and the environment the assistant can access. A workstation holding production credentials creates a different scope from an isolated working copy. Reviewing a software change should cover both the generated code and the operations the agent performed to produce it.

Reference documentation: MAI-2026-003.

Reliability and regulatory response

On 30 April 2026, Italian authority AGCM announced the closure of its Mistral investigation with commitments to improve information about fabricated or incorrect answers, without finding an infringement. Reliability notices are part of using the service, rather than evidence that errors are absent.

In prevention work, verification should focus on claims that change an action: a date, an owner, an instruction or a source’s scope. Record “not stated” when the material cannot answer a question. Asking the same assistant to confirm its answer may improve the writing, but does not provide independent evidence.

Reference documentation: AGCM: compromisos sobre alucinaciones.

Relationship to occupational safety and health

The following uses are Sabentis proposals based on the capabilities described. They do not represent an implementation or measured results.

  • Preparing meetings: organise previous agreements and unresolved questions with references. The value lies in locating the issues before the team meets, without turning a summary into approval.
  • Updating materials: compare a revised instruction with its predecessor and suggest which training passages need review. The responsible person decides whether a technical change affects actual work.
  • Exploring public information: gather institutional sources for a campaign and separate general recommendations from criteria applicable to the particular workplace. The selection should state audience, date and limitations.

Measuring pages produced is insufficient. To assess a pilot, count how many references are retrieved correctly, how many omissions review finds and how much effort is required to reach a usable document. It also matters whether the format helps people understand uncertainty instead of hiding it.

Reference documentation: Get started with Work · Files and Canvas · Search the web.

Hypothetical case: weekly prevention agreements

Suppose a team has three authorised sets of minutes and needs to prepare its weekly meeting. One mentions reviewing a walkway; another changes the deadline and a third reports partial work. The objective is to clarify the documented status, without declaring a risk closed.

  1. Prepare inputs. Save copies identified by date and version, removing unnecessary personal information. Identify the approved minutes when drafts also exist.
  2. Define extraction. Request a row per agreement with a reference, a summary of the original wording, the explicitly named owner, deadline and evidence of progress. Missing fields must remain “not stated”.
  3. Separate contradictions. Show the changed date with both references. Partial work is not interpreted as closure, and a suggestion does not become an accepted commitment.
  4. Check against the minutes. A person verifies all agreements to be communicated, including those the assistant omitted. Negations, names and dates receive particular attention.
  5. Prepare two outputs. One document contains confirmed agreements; another contains questions for the meeting. The team subsequently decides what to update in its management system and what to communicate.

The expected output is a traceable draft. Pilot acceptance could require a locatable reference for every selected agreement, no invented dates and visible contradictions. These criteria are a proposed verification method, rather than results Sabentis has measured in Vibe.

Reference documentation: Files and Canvas · Reuse work with Skills · Safety and approvals.

When to choose it and complementary tools

Vibe merits evaluation when combining research, documents and programming within one tool family is useful. An initial trial should represent ordinary work: a difficult document, a table containing exceptions and an output requiring review. Compare verifiable quality and review time before judging the fluency of a demonstration.

Microsoft Excel supports continued review of a workbook and its formulas. NotebookLM provides another approach to working with a document collection. The choice depends on where sources live, who will review them and which output must be retained. If formal tracking is needed, the organisation should maintain a register with responsibilities, permissions and verifiable status beyond the conversation alone.

Reference documentation: Get started with Work · Spreadsheets · Vibe Code.

Frequently asked questions

Are Le Chat and Vibe separate product entries?

This catalogue explains the assistant’s continuity and name change alongside its coding experience. Two identical entries would obscure that evolution.

Are summarised minutes equivalent to approved minutes?

No. Approval belongs to the team’s process. A summary can help locate agreements, but it must retain its relationship to the authoritative document.

Is a script that works once ready for real use?

Test the relevant cases: missing data, invalid dates, legitimate duplicates and inputs outside the expected format. A correct result on a simple sample does not cover those situations.

Reference documentation: Vibe overview · Vibe Code · Files and Canvas.

Request a Demo

Discover all that Sabentis can do for your organization.

Try Sabentis

request a demo
stars 5
GetApp Software Advice Capterra