What it does

Mistral Vibe is a command-line coding assistant that lets you talk to your codebase in natural language. It reads files, runs shell commands, searches with grep, and edits code — all from an interactive chat loop. You can also run it programmatically with --prompt for scripting, or even dictate commands via an experimental voice mode.

The interesting bit

The agent profiles and subagent delegation. Vibe ships with built-in personas like plan (read-only explorer) and accept-edits (auto-approves file changes), plus a task tool that spawns independent subagents to prevent context overload. It’s a rare CLI tool that acknowledges “vibe coding” might actually need guardrails — hence the trust folder system and per-tool approval tiers.

Key highlights

• Built-in agents: default, plan, accept-edits, auto-approve, plus custom agent configs via TOML
• Subagent delegation: Offload exploration or analysis to parallel task agents without cluttering your main session
• Programmatic mode: --prompt with cost caps (--max-price, --max-tokens) and JSON/streaming output for CI pipelines
• Interactive tooling: @file autocompletion, slash commands, persistent history, and a stateful bash terminal
• Safety layers: Trust folder confirmation, tool execution approval, and a .vibe directory-based project scoping

Caveats

• Windows support exists but is unofficial; the README targets UNIX environments
• Voice mode is explicitly marked experimental and may change
• Programmatic mode silently ignores your default_agent config and falls back to auto-approve — a gotcha for safety-conscious scripting

Verdict

Worth a spin if you want Claude Code-style agent interaction without leaving your terminal, especially if you already use Mistral’s API. Skip it if you need first-class Windows support or a GUI-first workflow.