Mistral Vibe
Minimal CLI coding agent by Mistral that lets you chat with your codebase and get things done.
Evidence shows: tool permissions are configurable (e.g., `[tools.bash] permission`), agents like `default`, `plan`, `accept-edits`, `auto-approve` are provided, default requires approval, supports `--auto-approve`/`--yolo`. Trust folder system requires user confirmation. API keys stored in `~/.vibe/.env`, environment variables supported. Dependencies pinned, CI uses pinned container digests. External effects include shell commands and file writes, but require approval. Rollback not explicit, only session management mentioned. Source attribution: publisher unverified, but repository has clear authors and license. Deductions: data flow transparency insufficient, no detailed telemetry or data collection; rollback mechanism not explicit; source attribution based on metadata only.
Evidence shows: README and config docs consistent, tool naming clear. Dependency list complete, versions pinned, CI tests cover. Failure messages: `/retry` command for backend errors, but no detailed error handling. Deductions: no troubleshooting guide, specific failure messages not shown.
Evidence shows: targeted at developers, supports multiple scenarios (interactive, programmatic, subagents). Capability boundaries via tool enable/disable, agent config. Trigger precision: `--agent`, `--enabled-tools` etc. Environment fit: Linux, macOS, Windows, but Windows support limited. Deductions: no detailed limitations per environment.
Evidence shows: README structure clear, install instructions detailed (curl, uv, pip). Naming stable: `vibe` command and config file names consistent. Examples abundant, FAQ section present. Known limitations: Windows warning. License: Apache-2.0. Version in pyproject.toml, but no CHANGELOG. Maintenance responsibility: publisher unverified, but CI and release workflow present. Deductions: no CHANGELOG, maintenance responsibility unclear.
Evidence shows: output formats support text, json, streaming, easy integration. Marginal value: multiple agents and skills system enhance flexibility. Cost-benefit: `--max-price` limit, but no pricing info. Deductions: cost-benefit not fully assessed.
Evidence shows: README feature descriptions consistent with code structure, but no implementation details. Cross-source corroboration: CI tests and PyPI link, but not independently verified. Fact-inference separation: warnings and descriptions distinguished, but some features may be inferred. Deductions: lack of traceable test results and independent verification.
- Publisher identity unverified, treat source with caution.
- Data collection and usage details incomplete, need further understanding of telemetry and privacy.
- Rollback mechanism not explicit, recommend backup before operations.
- Windows support limited, may affect cross-platform use.
- Dependencies pinned, but monitor security updates.
What does this agent do, and when should you use it?
Mistral Vibe is an open-source command-line coding assistant powered by Mistral's models. It provides an interactive chat interface that can read, write, and edit files, execute shell commands, search code with grep (including ripgrep), manage a todo list, delegate tasks to subagents, and ask the user clarifying questions. Configuration is done through a config.toml file, allowing customization of models, system prompts, tool permissions, and MCP servers. It includes a trust folder system for safety, multiple built-in agent profiles (default, plan, accept-edits, auto-approve), voice input (experimental), and session resume. The tool supports the Agent Skills specification for extensibility. It primarily targets UNIX environments; Windows support is partial.
Mistral Vibe runs as a CLI (command vibe) that starts an interactive session. It can read project files, use the grep tool (with ripgrep support) for code search, execute shell commands via bash, edit files with write_file and edit, maintain a todo list, and ask the user questions with ask_user_question. The task tool delegates work to subagents like explore. It integrates MCP servers for external tools. Programmatic mode uses --prompt for one-off runs, with options like --max-turns, --max-price, --max-tokens, and output formats text, json, and streaming. Voice mode is toggled with /voice and records via Ctrl+R.
- A developer wants to find all TODO comments in a project; they use Vibe's grep tool and then edit files.
- A user needs to refactor a function; they describe the task in natural language and Vibe performs the edits.
- In a large codebase, a user delegates exploration to the explore subagent while working on something else.
- A security-conscious user uses the plan agent for read-only exploration to avoid unintended changes.
- A developer automates a refactoring task using
vibe --promptwith budget limits and JSON output.
What are this agent's strengths and limitations?
- Ultra-lightweight CLI that starts fast and fits terminal workflows.
- Built-in agent profiles (plan, auto-approve) give fine-grained control over tool approval.
- MCP server support extends functionality to external tools and data sources.
- Voice input, session resume, and budget controls for programmatic use.
- Locked to Mistral API; no support for other model providers.
- Windows support is partial; official focus is UNIX.
- Requires Python 3.12+ and uv/pip; manual install steps.
- Voice mode is experimental and may be unstable.
How do you install or deploy this agent?
One-line install on Linux/macOS: curl -LsSf https://mistral.ai/vibe/install.sh | bash. Or use uv: uv tool install mistral-vibe. Or pip: pip install mistral-vibe. For Windows, install uv first, then use uv; official support is partial.
How do you use this agent?
Run vibe in your project root. On first run, it prompts for a Mistral API key (saved to ~/.vibe/.env) or configure with vibe --setup. Interactively type natural language commands, use @ to reference files, ! for shell commands, /help for commands. Programmatic: vibe --prompt "Refactor the main function" --auto-approve --output json.