By Mistral AILearning Objectives
- Understand what Mistral Vibe CLI offers and how custom subagents work
- Compare Vibe CLI with other terminal coding agents (Claude Code, Gemini CLI, Aider)
- Evaluate when Vibe CLI is the right choice for your workflow
What Is Mistral Vibe CLI?
Mistral Vibe CLI is the terminal entry point to the unified Mistral Vibe agent platform (see section-6-108). It is open source, powered by Devstral 2 (123 billion dense, 72.2% SWE-bench Verified) for local sessions, and runs on Mistral Medium 3.5 when launching remote cloud-sandbox agents.
The CLI sits alongside three other Vibe surfaces — the web UI at chat.mistral.ai (Work mode) and code.mistral.ai (Code mode), the Vibe VS Code extension, and the iOS and Android apps. Use the CLI when you want to drive Vibe from a terminal session without leaving your shell, scripted pipelines, or CI environment.
What sets the CLI surface apart from other terminal agents is custom subagents — define specialized agents via TOML configuration files, each with its own tools, context, and personality. This composability lets teams build domain-specific agents that chain together for complex workflows.
✅Tip
Get started: Install from github.com/mistralai/mistral-vibe — open source, on-premise capable
Core Capabilities
Custom Subagents (TOML Configuration)
Define specialized sub-agents in simple TOML files:
[agent.database-reviewer]
description = "Reviews database migrations for safety"
tools = ["file_read", "sql_explain"]
context = ["supabase/schema.sql", "docs/db-conventions.md"]
Each subagent has its own scope, tools, and context — preventing information leakage between tasks and enabling focused, high-quality work on specific domains.
Multi-Choice Clarifications
When a task is ambiguous, Vibe CLI presents structured options rather than guessing:
- "Did you mean: (A) refactor the auth middleware, (B) add new auth routes, or (C) update the auth tests?"
- This reduces wasted work from misunderstood prompts
Slash-Command Skills
Define project-specific skills as reusable commands — similar to Claude Code's SKILL.md pattern:
/deploy-staging— run your specific deployment workflow/add-api-route— scaffold a new API route following your project's patterns/run-migration— execute database migration with safety checks
MCP Support
Full MCP integration for connecting to external services — databases, deployment platforms, and APIs.
Remote Agents
Launch async cloud coding agents from the terminal with vibe remote run. Remote agents run independently of your local session in isolated sandboxes — the same Code mode infrastructure used by the Vibe web UI and VS Code extension. Use them for refactors, test generation across many files, scheduled maintenance jobs, and long-running migrations that would otherwise tie up the local session.
Remote agents run on the Mistral Medium 3.5 backend (128 billion-parameter dense, 256K context, 77.6% on SWE-Bench Verified and 91.4 on τ³-Telecom, with self-hosted versions running on as few as 4 GPUs). Local Vibe CLI continues to use Devstral 2 by default; the remote-agent path is opt-in. See section-6-199 for full Mistral Medium 3.5 specs.
Pricing
- Free and open source — on-premise deployment available
- Uses Devstral 2 via Mistral API (usage-based pricing) or self-hosted models
Strengths
- Custom subagents: Define domain-specific agents in TOML — unique composability among CLI tools
- Open source: Full source code available; on-premise deployment for air-gapped environments
- Devstral 2 powered: 72.2% SWE-bench Verified — strong coding model with 256K context
- Multi-choice clarifications: Reduces wasted work from ambiguous prompts
- Zed extension: Available as a Zed IDE extension in addition to standalone CLI
- Devstral Small 2 (24 billion, Apache 2.0) available for local deployment on consumer hardware
Limitations & Considerations
- Mistral ecosystem: Best with Mistral models; less tested with other providers
- Smaller community: Fewer users and integrations compared to Claude Code or GitHub Copilot
- Devstral 2 limits: 72.2% SWE-bench is strong but trails Claude Opus (80.8%) and Gemini 3.1 Pro (80.6%)
Best Use Cases
| Task | Why Vibe CLI |
|---|---|
| On-premise development | Open source + self-hostable — no code leaves your infrastructure |
| Domain-specific workflows | Custom subagents in TOML enable specialized agents per domain |
| Mistral ecosystem | Native integration with Devstral 2 and Mistral models |
| Zed users | Available as a Zed IDE extension — no separate installation |
When to choose alternatives:
- Maximum coding capability → Claude Code (Opus 4.7, 80.8% SWE-bench, 1 million context)
- Free daily usage → Gemini CLI (generous free tier with Google account)
- GitHub workflow integration → GitHub Copilot CLI (Issues, PRs, Actions)
- Model flexibility → Aider (works with any LLM, automatic git commits)
Key Takeaways
- Mistral Vibe CLI is the terminal entry point to the unified Mistral Vibe agent platform — sitting alongside the Vibe web UI, mobile apps, and VS Code extension
- Brings customizable subagents (TOML config) to terminal-based coding — unique composability for domain-specific workflows
- Powered by Devstral 2 (72.2% SWE-bench) for local sessions with on-premise deployment capability for air-gapped environments
- Vibe remote agents launchable from the CLI run on Mistral Medium 3.5 (77.6% SWE-Bench Verified) for parallel cloud-sandbox work
- Multi-choice clarifications reduce wasted work from ambiguous prompts
- Best suited for Mistral ecosystem teams, on-premise requirements, and developers who want composable agent architectures driven from a terminal
