Mistral released Vibe Desktop on November 17, the first native desktop application for its Vibe agent, the product rebranded from Le Chat earlier this year. The app is available for macOS and Windows with a Linux version planned, and it merges the local capabilities of Vibe CLI with the remote sessions of Vibe Code Web in a single window. Until now the agent lived in a web app, a terminal and a phone.
The core design is parallelism. Users can hand the agent several tasks at once and watch them run side by side, with Git worktrees keeping each parallel task in the same repository isolated in its own working directory so edits do not pile into one checkout. Mistral is careful to distinguish isolation of files from isolation of conversations: two chats opened against the same worktree still share its files, and dependencies or ports may need separate setup.
Execution happens in one of two environments. Local Code sessions run against the repository checkout, shell and toolchain on the user's machine. Cloud sessions clone a GitHub repository into a managed sandbox, with documented limits of 24 hours per session and up to three hours waiting for a reply. When a remote session ends, the sandbox and its build artifacts are deleted; branches, commits and pull requests already pushed to GitHub survive, but anything left unpushed does not. Local sessions share the CLI's core agent engine and session logs, though tools and prompts can differ between interfaces.
Work mode, Mistral's general productivity surface, can now take local folders as context, including report directories in different locations, and shares conversations, projects, skills, connected tools and scheduled tasks with the web app. Mistral's own documentation distinguishes adding a directory from actually reading every file in it, and sensitive Work actions trigger approval prompts. Delegated subagents get read-only views of their instructions and findings and report back through the main conversation.
The review surface is where the desktop form earns its keep: a unified or split diff view, pull request titles, numbers and status in an integrated browser, and agent summaries kept separately inspectable from the actual command output, with a stop control distinct from the one that kills background processes. Pricing follows the existing plans, with limited sessions on Free, Pro at $14.99 a month and Team at $24.99 per user, subject to fair-use caps Mistral does not publish in full.
What is missing is as telling as what shipped. Mistral does not mention any computer-use capability, while Anthropic's Claude desktop app has been opening applications, driving browsers and filling spreadsheets since the spring. On models, Mistral's October Large 4 release still trails Gemini 4 Carbon and Claude Opus 5.5 in independent comparisons. The French company's economics rest on enterprise and government customers running its models via API or on-premise, which is where a tidy desktop harness for regulated environments actually sells.
The desktop app race is really about where agents live when they stop being chat windows. Anthropic, OpenAI and Microsoft all shipped desktop surfaces this year, and each is betting that the operating layer of work is a window that drives other windows. Mistral's entry is late and narrower, but it is aimed at the one buyer segment, European enterprises and public agencies, where data residency and auditability decide procurement more than benchmark scores do.
Comments (0)
Log in to join the discussion
Log InNo comments yet