Mistral Vibe

// Overview

Mistral Vibe is Mistral's open-source command-line coding assistant. It puts a conversational agent in front of your codebase: you describe what you want in natural language and it explores, edits and runs things through a tool set covering file reads, writes and patches, managed shell sessions, recursive `grep` search (using ripgrep when available), a todo list it maintains as it works, an `ask_user_question` tool for when it needs your input, and a `task` tool that delegates work to subagents running in parallel.

Permissions are handled by picking an agent profile rather than answering the same prompt repeatedly. `ask` requires approval for every tool execution, `plan` is read-only and auto-approves safe tools like `grep` and `read` for exploration, `accept-edits` is the default and auto-approves file edits only, and `auto-approve` approves everything. Pass `--agent` to switch, or set `default_agent` in `config.toml`. On top of that sits a trust-folder system so an agent's freedom depends on where it is running.

Vibe scans a project's file structure and Git status to build context automatically, and the CLI itself is built for long sessions — slash-command and `@` path autocompletion, persistent history, themes, and image attachments through `@` mentions that are sent as native multimodal content to vision-capable models. Everything else is configuration: models, providers, tool permissions, OpenTelemetry tracing, custom system prompts, custom agent definitions, MCP servers and session handling all live in a `config.toml`. Skills add custom slash commands. Mistral officially supports UNIX environments; Windows works but is not the target.

// What it does

• Terminal agent with read/write/edit, managed shell sessions, ripgrep-backed search, a todo tool and interactive user questions
• Subagent delegation through a `task` tool for parallel work
• Four built-in agent profiles — ask, plan, accept-edits, auto-approve — plus a trust-folder system, instead of per-call approval fatigue
• Project-aware context from your file structure and Git status, gathered automatically
• Skills system that adds custom slash commands, and MCP server configuration for external tools
• Configured through a single config.toml: models, providers, tool permissions, custom system prompts, OpenTelemetry tracing and sessions

// Getting started

Vibe needs Python 3.12+. The one-line installer is the recommended path on Linux and macOS; uv and pip both work everywhere.

curl -LsSf https://mistral.ai/vibe/install.sh | bash

# or
uv tool install mistral-vibe

# or
pip install mistral-vibe

cd /path/to/your/project
vibe

vibe --agent plan

default_agent = "plan"

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Work through a codebase from the terminal with a Mistral-model agent instead of an IDE plugin
• Explore an unfamiliar repository safely in read-only plan mode before letting anything write
• Refactor across many files under the default accept-edits profile, with shell commands still gated
• Fan a large task out to subagents, or wire in MCP servers and custom skills for project-specific workflows

// How Mistral Vibe compares

Mistral Vibe alongside other open-source autonomous coding agents tools AI/TLDR tracks, ranked by GitHub stars.