Back to Gap Map Product / UX / Orchestration & agents

Mistral Vibe

Mistral AI
open source / Overall score: 4.0(strong)

Mistral Vibe is a terminal coding agent that reads and edits files, runs shell commands, searches code and delegates work to subagents. It runs interactively, as a non-interactive command for scripts, or inside editors through the Agent Client Protocol, and extends through MCP servers, skills and hooks. Mistral's hosted models are the default, though local models and other compatible API keys also work. Mistral AI develops it.

Mistral's site sells this as Vibe for code, inside the Vibe product that replaced Le Chat, which keeps its own entry. The optional Rust terminal client's Cargo manifest says "Proprietary", but the repository's only license file is Apache-2.0, and the entry reads that.

Openness

5 medium confidence
5.0
license
Apache-2.0(OSI)
source
public
core features withheld
no — the 2.26.1 source distribution carries one Apache-2.0 LICENSE and no ee/, enterprise/, commercial/ or proprietary/ path, Mistral's launch post says the CLI is released under Apache 2.0, and its docs tie paid plans to model usage and rate limits rather than to CLI features

Mistral Vibe is Apache-2.0 licensed, and its public repository holds the Python agent, the native harness and the optional Rust terminal client, so the published wheels build from source. Mistral's paid plans buy higher model usage and rate limits rather than CLI features, and the CLI also runs offline with local models.

Adoption

4 high confidence
4.0

Adoption is the trailing month of PyPI downloads of mistral-vibe, the one package the CLI ships as. The README's uv and pip instructions and Mistral's one-line installer all install that package, so the count covers every documented install route.

Capability

4 medium confidence
4.0

Mistral Vibe covers a full terminal coding agent with subagents, a read-only planning profile, approval controls, MCP and editor use through the Agent Client Protocol. It sits a step below OpenHands, which runs agents in a sandboxed runtime and publishes benchmark results, while Vibe runs on the user's machine and has no benchmark of its own.

  • https://docs.mistral.ai/vibe/code/cli/install-setup recorded 2026-10-19

    The setup docs say "The CLI also works fully offline with local models, or against any compatible API key you provide."

  • https://mistral.ai/news/devstral-2-vibe-cli recorded 2026-10-19

    The launch post reports SWE-bench Verified scores of 72.2% for Devstral 2 and 68.0% for Devstral Small 2, and human evaluations "with tasks scaffolded through Cline"; it gives no benchmark figure for the CLI.

  • https://raw.githubusercontent.com/mistralai/mistral-vibe/main/README.md recorded 2026-10-19

    The README lists "Delegate tasks to subagents for parallel work (`task`)", a "`plan`" profile described as "Read-only agent for exploration and planning", MCP server configuration, hooks, skills and Agent Client Protocol editor support; its sandbox option is marked "Experimental".

Verified 2026-10-19, except some scores, still awaiting re-confirmation