Mistral Vibe

In the cloud, coding agents work through the Mistral GitHub App with the user's token to clone repositories, branch, commit, push and open pull requests, and the CLI connects to remote MCP servers with OAuth, header or API key auth and to local stdio servers, alongside Mistral managed connectors. The work agent reaches Gmail, Google Calendar, Outlook, SharePoint, Box, Slack, Notion, Linear, Atlassian, GitHub and Stripe through OAuth, plus a directory of MCP connectors and custom MCP servers that admins add.

Sourcedocs.mistral.ai/vibe/work/connectors.mdread 2026-11-08

The CLI hands parts of a task to subagents through a task tool, with a built in read only explore subagent and custom subagents a team defines, and the agent's plan can be edited before it runs. Remote agents run in parallel, each in its own isolated cloud sandbox, for up to 24 hours with a three hour idle timeout. The work agent carries multistep tasks through a todo panel, asks clarifying questions as it goes and can be stopped at any point.

Sourcedocs.mistral.ai/vibe/code/cli/agents.mdread 2026-11-08

Libraries hold the customer's documents as persistent knowledge bases with real time search and footnoted sources, up to 100 files per upload and 100 MB per file, stored in Mistral's European cloud, private by default and shareable with viewers or collaborators, and Team adds knowledge connectors. For code, the agent works from the project path, branch, recent commits, grep and AGENTS.md files on each run, with no persistent index of the codebase.

Sourcedocs.mistral.ai/vibe/work/libraries.mdread 2026-11-08

The CLI runs in ask, accept edits (the default, which approves file edits on its own), smart approve or auto approve modes, with per tool always, ask or never rules, trusted folders, prompts before leaving the working directory or touching environment files, and hooks before and after tools that can block commands, and admins can forbid auto approve. Remote agents deliver pull requests for review under the repository's own branch protections, and the work agent asks Continue, Always allow or Decline before acting.

Sourcedocs.mistral.ai/vibe/code/safety-approvals-permissions.mdread 2026-11-08

Enterprise signs in through SAML 2.0, and roles span organization roles for members, billing and admins and workspace roles such as Developer, Mistral Vibe Code User, Contributor, Admin and Observability Viewer. Managed configuration pushes allowed models, connector blocks and approval limits to the CLI and VS Code. Mistral states it complies with the SOC 2 Type II and ISO 27001 and 27701 frameworks, Enterprise is opted out of model training by default, and Team admins can turn training off for their organization.

Sourcedocs.mistral.ai/admin/identity-access/roles-permissions.mdread 2026-11-08

Pro and Team keep run history, tool calls, code changes, logs and audit trails for remote agents, and each web session's activity stream serves as its audit log. The CLI writes local logs, exports a run to JSON and sends opt in OpenTelemetry traces with redaction, Enterprise audit logs record when conversations are created and deleted, and the Enterprise admin API reports usage analytics for coding agents.

Sourcedocs.mistral.ai/admin/monitor-comply/audit-logs/overview.mdread 2026-11-08

The work agent's Knowledge Base, in public preview, saves facts when asked and on its own, shows each read and write in the chat, and lets users list and edit entries and delete one, a topic or everything, export them from settings or turn the feature off, while admins can disable it for an organization and sensitive data stays off by default. Workspace chat retention can be set from 30 days to never, temporary chats skip memory, and CLI and VS Code sessions resume across both.

Sourcedocs.mistral.ai/vibe/work/knowledge.mdread 2026-11-08

Data is hosted in the European Union by default, with a US endpoint available, and Enterprise deploys on premises, in a private cloud or on Mistral Cloud, including a self hosted option installed with Helm. The CLI runs on the developer's machine and can run fully offline against local models, contacting no Mistral server once telemetry and update checks are turned off.

Sourcedocs.mistral.ai/vibe/code/cli/offline-models.mdread 2026-11-08

The work agent ships twelve built in skills, among them deep research, data analysis, meeting prep and a skill creator, following the Agent Skills standard, and admins can switch on workspace skills for everyone. The CLI includes built in agents such as a lean mode and a plugin format, labeled experimental, with a built in Vibe plugin, and teams add their own skills and subagents alongside these.

Sourcedocs.mistral.ai/vibe/work/skills.mdread 2026-11-08

Coding work starts from the CLI, interactively or headless with a prompt flag, from the VS Code extension, from JetBrains, Zed and other editors over the Agent Client Protocol, from the Code tab on the web, or by sending a CLI task to the cloud. The CLI's loop command runs cron schedules while Vibe is open and idle, and the work agent runs scheduled tasks once, daily, weekly, monthly or yearly, in public preview and up to five on Free. Starts from Slack and from GitHub events are still rolling out or planned.

Sourcedocs.mistral.ai/vibe/work/scheduled-tasks.mdread 2026-11-08

In the CLI, developers pick the model with a command, bring their own API key or point Vibe at any OpenAI compatible local model, with Devstral Small 2 or Mistral Small 4 suggested for local use. Remote coding sessions and the work agent run on Mistral's own models, with Mistral Medium 3.5 as the default and only a fast or think toggle beyond it, and Enterprise can fine tune models through Mistral Forge.

Sourcedocs.mistral.ai/vibe/code/cli/offline-models.mdread 2026-11-08

Vibe runs as an Agent Client Protocol server listed in the ACP registry, so editors and other clients drive it programmatically, and the CLI runs headless with a prompt flag and JSON or streaming output for scripts and CI. An agent socket and a local run server are experimental, and Enterprise adds an admin API. Vibe has no MCP server of its own, and Mistral's API and agents SDK belong to its developer platform.

Sourcegithub.com/mistralai/mistral-viberead 2026-11-08

Headless runs return exit codes and export a JSON record of what the agent did, so a team can score runs in its own pipeline, and remote agents hand back pull requests that the repository's own checks and reviewers test. Vibe itself has no evaluation harness or scored test set for its agents, and Mistral's evaluation tools sit in its developer platform in public preview.

Sourcegithub.com/mistralai/mistral-viberead 2026-11-08

The work agent searches the web and opens URLs to read pages, and remote coding sandboxes have outbound internet access, but no Vibe agent drives a browser, a desktop or another computer through its interface. Web content reaches the agent as fetched text, and coding agents work through the shell, files and Git in the terminal or in their cloud sandbox.

Sourcedocs.mistral.ai/vibe/work/safety-and-approvals.mdread 2026-11-08