Mistral Vibe: Free CLI Coder with Medium 3.5

AICodeKing2026-05-14go watch the original →

the gist

Mistral Vibe runs as a terminal CLI agent using free Experiment API (rate-limited, may train models) with Mistral Medium 3.5 (128B params, 256k context, 77.6% SWE-Bench Verified) for repo tasks via @file refs, !shell cmds, /slash commands.

Free Setup and Installation

Users install Mistral Vibe on Mac/Linux (WSL for Windows) with Python 3.12+ via one-line curl, uv, or pip. They run vibe-setup in a project directory, paste a free Experiment API key from Mistral AI Studio (organization keys page, name it, set scope), which stores locally. This plan suits prototyping but warns requests may train Mistral models; avoid private code.

Paid Le Chat Pro ($14.99/mo) or Team ($24.99/user/mo) add full Vibe access, higher limits, privacy. API pricing for Medium 3.5 is $0.50 per million output tokens.

Core Commands and Workflow

Vibe reads repo structure, Git state, enables file reads/writes, patches, searches, shell cmds, tests, todo lists, approval gates. Key syntax: @filename autocompletes for file refs (e.g., @src/app.js), !cmd runs shell directly (e.g., !git status, !npm test), /help lists options, /model changes settings, /config updates, /clear resets, /teleport moves to cloud sandbox.

Non-interactive: vibe --prompt "analyze codebase" --max-turns 5 --max-price 0.1. Workflow: Prompt to inspect/plan first (e.g., "explain project structure, main entry points"), approve plan, implement small parts, run tests, review diffs.

Medium 3.5 Model and Remote Agents

Mistral Medium 3.5 (128B dense, open-weights modified MIT, 256k context, launched April 2026) supports configurable reasoning (low for quick, high for agentic), tool calls, structured outputs. Benchmarks: 77.6% SWE-Bench Verified. Powers local terminal, remote cloud sessions (async, diffs/PRs), Le Chat work mode for multi-tool tasks.

Remote: Start local, /teleport to cloud for long tasks (e.g., tests, CI fixes); integrates GitHub/Linear/Jira/Slack.

Practical Use Cases

Codebase mapping: "Explain architecture, find main routes, summarize DB layer."
Tests: "Add tests for user auth flow" (matches style, runs/fixes).
Refactors: "Extract validation logic, split giant file, separate data/UI."
Bugs: Paste error, inspect/reproduce/patch/rerun.
Docs: Generate README, setup steps, env vars.

Best for contained tasks; compare to Claude Code, Gemini CLI, Open Code, Qen Code.