Mistral Vibe: Open-Source CLI Coding Agent and Claude Code Alternative

Mistral Vibe is Mistral AI’s open-source command-line coding assistant.

Open a project in the terminal, describe a task in plain language, and ask it to trace a bug, explain part of the codebase, refactor a component, or prepare a review change.

It reads the project structure and Git status, accepts file references with @, and loads local instructions from AGENTS.md.

Start with the read-only planning profile when you need to understand a codebase, then use an editing profile after you have defined the work.

Mistral Vibe is designed as an alternative to Claude Code for developers who prefer Mistral’s model family and want to inspect the command-line client they run.

It is Apache-2.0 open source and supports project instructions, skills, MCP servers, hooks, permissions, and saved sessions.

Key Features

• CLI coding assistant: Run vibe inside a project to inspect files, apply patches, run commands, and track tasks from the terminal.
• VS Code support: The Mistral Vibe VS Code extension brings agentic code editing, repository context, and change approval into the editor.
• Project-aware context: Vibe reads project structure, Git status, mentioned paths, trusted folders, extra directories, and local instructions before it works on a repository.
• File and terminal tools: The CLI can read and edit files, search code with grep, run shell commands, manage a task list, ask questions, and delegate work.
• Built-in agents: The CLI ships with default, plan, accept-edits, and auto-approve profiles. Choose a profile with --agent or set a default for interactive sessions.
• Subagents: Vibe delegates work through the task tool. The built-in explore subagent is read-only and keeps codebase exploration out of the main session context.
• Interactive questions: The ask_user_question tool lets the agent stop mid-task and ask for a choice or a free-text answer before it continues.
• Skills and slash commands: Skills add reusable instructions, tools, slash commands, and specialized behavior. Vibe discovers them from project-level and user-level locations, including Agent Skills standard paths.
• Hooks: Define shell-based checks in hooks.toml to enforce repository rules, deny unsafe commands, add context, or trigger local automation around agent activity.
• MCP server support: You can configure Model Context Protocol (MCP) servers with http, streamable-http, or stdio transports, custom headers, API key environment variables, timeouts, and tool permissions.
• Session controls: Use --continue, --resume, --workdir, and --add-dir to continue work, resume a saved session, choose a working directory, or include additional directories.
• Safety controls: Tool approval, trusted folders, configurable permissions, and agent profiles separate read-only planning from edit-heavy sessions.

Mistral Vibe Pricing and Free Plan

Mistral offers a Free plan with limited Vibe access and limited coding sessions. You can try the CLI on a real project before paying.

The Pro plan costs $14.99 per month, excluding taxes. It includes full Vibe access for long-running tasks and all-day coding in the CLI, IDE, and web, subject to fair-use limits. It also raises allowances for messages, web searches, image generation, and support.

The Team plan costs $24.99 per user per month, excluding taxes. It adds a shared workspace, 30GB of storage per user, domain verification, and data export. Enterprise pricing covers private deployments, custom models, custom agents, custom workflows, audit logs, SAML SSO, and white-label options.

Verified students can get the Education plan at $5.99 per month, excluding taxes, with eligibility limits.

For API usage, Mistral bills per million input and output tokens. The pricing page lists Devstral 2, the agentic coding model, at $0.40 per million input tokens and $2.00 per million output tokens.

This API pricing is separate from the consumer plan framing and matters if you extend usage through pay-as-you-go credits or build directly on the API.

Use Cases

• Legacy system modernization: Ask the agent to refactor older code toward a newer architecture, then review the patch before you merge it.
• Cross-file bug fixes: Describe a bug that spans multiple files. The agent can search the repository, inspect related code, and propose a consistent fix.
• Dependency updates: Use the agent to find imports, update package references, and adjust code affected by changed APIs.
• Codebase exploration: Use the built-in plan agent or explore subagent to map an unfamiliar repository before you ask for edits.
• Automated documentation: Generate README updates, API notes, migration notes, or inline comments after the agent inspects the relevant files.
• Test generation: Ask Vibe Code to write unit or integration tests that match the patterns already used in the project.
• Repeatable engineering routines: Create skills for code review, release checks, security review, documentation, or repo-specific conventions.
• Pull request preparation: Ask the agent to inspect a completed change, run the relevant checks, summarize the diff, and flag files that need a human review.
• Image-informed coding tasks: Attach a screenshot or design reference with @ when a compatible vision model needs to inspect an interface or visual bug.
• VS Code tasks: Use the extension when you want to review agent changes and work with repository context inside the editor.

How to Use Mistral Vibe CLI

Installation

One-line install for Linux and macOS:

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

Using uv:

uv tool install mistral-vibe

Using pip:

pip install mistral-vibe

Windows: Vibe works on Windows, while Mistral officially targets UNIX environments. Install uv first:

powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

Open a new PowerShell window, then run uv tool install mistral-vibe.

Run vibe --help after installation to confirm that your shell can find the command before you open a repository.

Start a Project Session

Navigate to your project root and run vibe. On first launch, the CLI creates ~/.vibe/config.toml, asks for a Mistral API key if one is not configured, and saves credentials to ~/.vibe/.env. You can also run vibe --setup for separate setup.

You can set the API key with MISTRAL_API_KEY or a ~/.vibe/.env file. Environment variables take precedence over values saved in the dotenv file. General settings belong in config.toml, not the .env file.

Start from the repository root so Vibe can read the project structure and Git status. Keep API credentials out of the repository. Use the user-level ~/.vibe/.env file or an environment variable, and add any project-specific behavior through the project configuration.

Choose an Agent Profile

Vibe includes four primary agent profiles. default asks before tool execution. plan is read-only for exploration and planning. accept-edits approves file edits, and auto-approve approves tool calls. Select one for a session with vibe --agent plan, or set default_agent in config.toml.

Start with plan when you need to understand an unfamiliar codebase. Move to an editing profile only after you know the target files and expected outcome. Review each diff and test result before you commit generated changes.

Interactive Mode

Run vibe to start the interactive chat loop. Use Ctrl+J or Shift+Enter for multi-line prompts, type @ to reference files with path autocompletion, and prefix a command with ! to run it directly in your shell.

You can press Ctrl+G to edit input in an external editor, Ctrl+O to toggle tool output, Ctrl+T to toggle the todo list, and Shift+Tab to cycle through available agents.

You can also start directly with a prompt:

vibe "Refactor the main function in cli/main.py to be more modular."

Give the agent a concrete task and a boundary. For example, name the failing test, the files it may edit, and the command it should run after the change. Use @ to attach a file or image, and use ! only for commands you intend to run yourself outside the agent workflow.

Type /help when you need the available session commands. Use Ctrl+O to show or hide tool output and Ctrl+T to inspect the task list. Keep both visible during a long task, then review the command output before you accept an edit.

Review Project Trust

Vibe uses a trust folder system for directories that contain Vibe configuration. When you first run Vibe in a new directory with a .vibe subfolder, it may ask whether you trust that folder. Trusted folders are saved in ~/.vibe/trusted_folders.toml.

Automate a Task

Use programmatic mode for scripts, repeatable checks, and non-interactive tasks:

vibe --prompt "Analyze the codebase" --max-turns 5 --max-price 1.0 --max-tokens 50000 --output json

Important options include --max-turns, --max-price, --max-tokens, --enabled-tools, and --output. Output can be text, json, or streaming. Programmatic mode uses your configured default agent. Add --auto-approve or --yolo only when the script has clear tool limits, cost limits, and a safe workspace.

Add Project Skills

Skills keep reusable task instructions close to your work. Put a skill folder with a SKILL.md file in .vibe/skills/, .agents/skills/, or a configured skill path.

Use enabled_skills and disabled_skills in config.toml to control discovery. Keep skills from untrusted repositories out of the active paths until you have read their instructions and tool allowances.

Voice Mode

Voice mode is experimental. Use /voice to toggle dictation. Ctrl+R starts recording, any key stops recording, and Escape or Ctrl+C cancels recording.

Configure Vibe

Vibe looks for ./.vibe/config.toml first and then falls back to ~/.vibe/config.toml. The configuration file controls the active model, default agent, tool permissions, skill paths, MCP servers, prompts, session behavior, update behavior, telemetry, notifications, and UI preferences.

You can set VIBE_HOME to store configuration, API keys, agents, prompts, tools, and logs in another directory. This helps keep work and personal coding environments separate.

Add Project Rules with AGENTS.md

Mistral Vibe supports AGENTS.md files for instructions. A user-level file at ~/.vibe/AGENTS.md applies across projects, while project-level files are loaded from the current directory up to the trust root. Closer directories override more distant instructions.

You can also replace the system prompt with a Markdown file in ~/.vibe/prompts/ or a project-local .vibe/prompts/ directory, then set system_prompt_id. For compaction, set compaction_prompt_id to control how long sessions are summarized.

Control Agents and Tools

Create custom agent configurations in ~/.vibe/agents/ or a project-level agents directory. A custom agent can specify a model, prompt, disabled tools, and per-tool permission behavior. Use enabled_tools and disabled_tools with exact names, glob patterns, or re: regular expressions to control what the agent can use.

Connect MCP Servers

MCP servers are configured under mcp_servers. Vibe supports http, streamable-http, and stdio transports. You can set URLs, headers, API key environment variables, commands, arguments, environment variables, startup timeouts, and tool execution timeouts.

MCP tool names follow the pattern {server_name}_{tool_name}. For example, a fetch server tool can appear as fetch_server_get. You can assign tool permissions to MCP tools the same way you assign permissions to built-in tools.

For a hosted OAuth server, use /mcp add from the CLI and complete the sign-in flow. For a local or API-key server, add its connection details to config.toml. Set startup and tool timeouts when a server launches slowly or serves long-running requests.

Resume a Session

Use vibe --continue or vibe -c to continue the most recent saved session. Use vibe --resume SESSION_ID to resume a specific session. Session logging must be enabled for these features.

Use --workdir to choose a working directory and --add-dir to make additional directories available for a session. Added directories are implicitly trusted and can contribute their own instructions, configuration, skills, prompts, and agents.

Updates, Notifications, and Telemetry

Auto-update is enabled by default. Set enable_auto_update = false in config.toml to disable it. Notifications can alert you when the agent needs approval, asks a question, or completes a task. Set enable_notifications = false to disable them.

Use of Vibe can include collection and processing of usage data. You can disable telemetry by setting enable_telemetry = false in config.toml. If you work with private code, client projects, or regulated data, check this setting before you start a session.

Mistral Vibe vs Claude Code

Mistral Vibe and Claude Code solve the same basic problem: you describe a software task, the agent reads project context, edits files, runs tools, and returns changes for review.

The difference is in the model family, product surface, pricing, openness, and how much you care about remote coding sessions.

Pros

• Try the CLI with free sessions
• Keep project rules in local files
• Choose read-only or editing profiles
• Delegate repository exploration to subagents

Cons

• Free coding capacity is limited
• Requires Python 3.12 or newer
• Windows has secondary support
• You must review agent permissions

Alternatives and Related Resources

• 7 Best CLI AI Coding Agents (Open Source)
• MiMo Code: Free Open-Source AI Coding Agent with Persistent Memory
• Grok Build: SpaceXAI’s Open-source Terminal AI Coding Agent
• Most Popular Agent Skills on GitHub for Coding Agents

FAQs

Q: Does Mistral Vibe VS Code require a subscription?
A: The extension requires an active Vibe Pro or Team subscription, or API access through Mistral Studio.

Q: Can Vibe use my company certificate authority?
A: Yes. Set enable_system_trust_store = true in config.toml to use the operating system trust store. You can also provide extra trust anchors through SSL_CERT_FILE or SSL_CERT_DIR.

Q: Which hook event names are current in Vibe?
A: Use pre_tool, post_tool, and post_agent in hooks.toml.

Changelog

July 19 2026 update:

• Updated Windows installation guidance and current coding-plan details.