---
title: "Mistral Vibe"
description: "Mistral's CLI. Everything under ~/.vibe and trust per folder. MCP servers as TOML tables and a read-only mode down to read_file."
canonical_url: "https://harnesses.agntn.dev/harnesses/vibe"
---
# Mistral Vibe

> Mistral's CLI. Everything under ~/.vibe and trust per folder. MCP servers as TOML tables and a read-only mode down to read_file.

<harness-facts id="vibe">



</harness-facts>

Mistral Vibe is Mistral's coding CLI, and the binary is `vibe`. It installs as a Python tool, but the launcher just hands over to a bundled Rust binary. Neat.

Everything lives under `~/.vibe/`, and `VIBE_HOME` moves all of it. User config is `~/.vibe/config.toml`, a project adds `.vibe/config.toml`. A `VIBE_<KEY>` variable overrides one key. Hooks sit in `hooks.toml`, in both places.

Is your project folder trusted? If not, Vibe skips its config, its skills and its `AGENTS.md`. A headless run warns, then answers without them. Trusted folders go into `~/.vibe/trusted_folders.toml`, and `--trust` covers a single run. The registry won't add that flag for you.

`agents sync` links `~/.vibe/AGENTS.md` for your own instructions. Project ones are every `AGENTS.md` from the working directory up to the trusted root, the closest winning.

Skills come from `.vibe/skills/` and `.agents/skills/` in a project, `~/.vibe/skills/` and `~/.agents/skills/` at home. Skills sync owns `~/.vibe/skills/`, so after a full sync Vibe finds each skill twice and quietly keeps the first. Custom slash commands are skills too, so there's no prompt template folder.

Sessions get a directory each under `~/.vibe/logs/session/unified/<session-id>/`. Inside: `meta.json`, journal JSONL files and content chunks.

MCP servers are `[[mcp_servers]]` tables in `config.toml`, each one naming itself. `vibe mcp add` writes the user file. Transports are `stdio`, `http` and `streamable-http`, no `sse`. A name with a space or a dot gets rewritten on load, so the library refuses it before writing. Better than a server that answers to a name you never gave it.

`vibe --prompt=<text>` is the headless mode. The prompt stays glued to `--prompt=`, because `-p` reads a prompt that starts with a dash as a flag.

`--output json` prints the session messages as one JSON array at the end. `--model` takes a catalog alias like `mistral-medium-latest`, and a wrong one fails with the list of real ones. `--thinking` carries effort. An unknown word gets a warning, then no thinking level at all.

Runs use `default_agent`, which is `accept-edits` unless changed. Nobody answers approval prompts headless, so anything that would ask gets denied.

Read-only passes `--disabled-tools` with a pattern that hits everything but `read_file`. Disabled tools add up across config layers and agent profiles, so no profile can switch one back on.

`--enabled-tools` wasn't enough: a custom profile with its own list replaced it and wrote files. `run_typescript`, `search_tool_functions` and `wait` stay. That sandbox has no files, no `fetch` and no `process`. Bash, edits, subagents, skills and MCP are all denied.

There's no advisor mode. Switch every tool off and the model still sees the whole toolbox. It then spends its turns knocking on locked doors.

Every command Vibe starts gets `VIBE_SESSION=1`. That one variable is all `detectHarness` needs. A `.vibe` folder is the project marker.

<harness-sheet id="vibe">



</harness-sheet>


## Sitemap

See the full [sitemap](https://harnesses.agntn.dev/sitemap.md) for all pages.
