Mistral Vibe: the open-source terminal coding agent

What Mistral Vibe is (and how it got here)

December 9, 2025: Mistral releases the Vibe CLI source on GitHub under Apache 2.0 alongside Devstral 2, its coding-model family. It is a Python agent built on Pydantic and Rich/Textual — the same stack as OpenHands, against Gemini CLI's TypeScript and Codex CLI's Rust. The design decision that matters: every core tool and the system prompt live as Markdown files in core/prompts/, so you can read exactly how the agent is instructed without decompiling anything.

January 27, 2026: Mistral Vibe 2.0 ships — custom subagents, multi-choice clarifications, slash-command skills, unified agent modes, and continuous auto-updates. May 28, 2026: Vibe becomes Mistral's unified agent for work and code. Le Chat is renamed Vibe (conversations and plans carry over), and a new Code Mode runs remote coding agents from the web app in isolated sandboxes — sessions persist while your machine is off, run in parallel, and /teleport moves a live session between terminal and cloud with history and approvals intact. A VS Code extension runs the same harness as the CLI; JetBrains and Zed connect over ACP.

The release cadence is fast and observable: the public changelog shows 2.12.x in late May, 2.15.0 in mid-June, and 2.23.3 by early August 2026 — MCP server management (vibe mcp add), AGENTS.md support at user, project, and parent-folder levels, and voice mode with real-time transcription all shipped in that window. Plan for the agent you reviewed last month to not be the agent you run today.

What version 2.0 actually added

Under it all sits the same terminal-agent grammar the rest of the ecosystem standardized in 2025–26: reference files with shortcuts, run shell commands, full-repo context, and tab completions trained on your codebase via Codestral. Connectors pull context from GitHub, GitLab, Jira, and private codebases; remote MCP servers attach with vibe mcp add without editing config files.

Install and first run

Two installers ship on the official page — a curl script or uv, which is the cleaner path if you already manage Python tooling:

curl -LsSf https://mistral.ai/vibe/install.sh | bash
# or, with uv:
uv tool install mistral-vibe

After auth, drop an AGENTS.md at the repo root — Vibe reads it from ~/.vibe/ (user-level), the project root, subfolders, and parent folders, so monorepo teams can scope instructions per package. Extend with skills via slash commands and remote MCP servers via vibe mcp add. Auto-updates are on by default; vibe update forces one.

Pricing: subscription-first, API-optional

Vibe is bundled into chat subscriptions rather than sold as a per-seat agent license, with pay-as-you-go beyond plan limits and a bring-your-own-key escape hatch. Verified against Mistral's January 2026 release post and product pages.

The subscription-first model is the differentiator: Claude Code and Codex CLI bill through API credits or per-seat plans, while Vibe's Pro plan is explicitly sized for sustained daily dev work. Verify current numbers on mistral.ai before budgeting — this market reprices quarterly.

Mistral Vibe vs Claude Code

Both are terminal-native coding agents with subagents, skills, and MCP support. The differences are license, stack transparency, model access, and where the bill lands.

Honest read: Claude Code still has the stronger model for hard agentic work and the larger plugin ecosystem. Vibe wins when you want an auditable agent — open prompts, open source, your own key — or a flat subscription that covers full-time use. If you already standardized on Anthropic, switching costs outweigh the license difference; if you are model-flexible or budget-capped, Vibe is the strongest open option in the terminal.

Where VibeFuse fits: an open harness with a marketplace

Mistral Vibe proves the direction — agents as products, skills as an open standard — but its skills and subagents stay inside Mistral's stack. VibeFuse is the open counterpart: the first free widget-based AI harness, with an open-source marketplace where creators sell widgets, skills, and styling packs at an 80% revenue share. The same slash-command skill you would write for Vibe becomes a product you can publish and get paid for — and the harness runs locally and offline on Windows, in any app, with no vendor lock on your workflow.

If you are evaluating terminal agents in 2026, the practical stack is complementary: Vibe (or Claude Code, or Codex) for the code-generation loop; VibeFuse for the verification and customization layer around your apps — review inside the loop, widgets from the marketplace, and local processing that keeps your code and audio on your machine.

Related reading

Mistral Vibe — FAQ