A terminal coding agent that delegates grunt work and caps your API bill

What it does Mistral Vibe is a terminal-based coding assistant that connects Mistral’s models to your local project. You chat with it in natural language, and it can read files, run shell commands, search code with grep, and modify your codebase through a set of built-in tools. It scans your project structure and Git status automatically to keep context relevant.

The interesting bit Instead of a single monolithic agent, Vibe lets you spin up read-only explore subagents to analyze architecture in parallel, keeping the main conversation uncluttered. It also treats caution as a feature: built-in profiles like plan and accept-edits auto-approve only specific tool categories, and a trust-folder system warns you before running inside sensitive directories.

Key highlights

• Built-in toolset covers file I/O, shell execution, grep/ripgrep search, and interactive user questions with multiple-choice prompts.
• Agent profiles (default, plan, accept-edits, auto-approve) control which tools require manual approval.
• Subagent delegation via a task tool offloads exploration or analysis to parallel workers.
• Programmatic mode supports cost caps (--max-price), token budgets (--max-tokens), and JSON or streaming output for scripting.
• Experimental voice mode for dictating prompts, plus slash commands and a skills system for custom commands.

Caveats

• Official support targets UNIX; Windows compatibility is acknowledged but not guaranteed.
• Voice mode is explicitly marked experimental and may change.
• Requires a modern terminal emulator (WezTerm, Alacritty, Ghostty, or Kitty are recommended); older terminals may render the UI poorly.

Verdict Developers who want a scriptable, safety-conscious CLI coding assistant with Mistral models should look here; if you need deep IDE integration or guaranteed Windows support, this is not your tool yet.