What it is
Mistral’s open-source coding CLI agent. The most useful thing to study here is what a reference agent loop looks like at minimal scale, before any production hardening.
What’s worth studying
- A clean while-loop with OpenAI-style tool calling. The shape is the diagram from agent-loop with no extra layers. Read this when you want to remember what an agent loop actually is, before all the cache-control and condenser machinery.
- Middleware-based dispatch. Tool calls flow through a middleware chain (validate → enforce policy → execute → record). Easy to extend, easy to test individually. Worth borrowing for your own dispatch when you’re past the prototype phase.
- Skills system via SKILL.md frontmatter. Same idiom as Strix and Claude Code — versionable, contributor-friendly. See skills-as-md.
The cost-side stuff (caching, compaction) is intentionally absent, which is part of why the codebase is readable. When you scale, you’ll add those layers; this is the substrate.
Drill-down
The full per-doc analysis lives below — these are the original numbered analyses, rendered as styled HTML. Pick a section to study deeper.
- Mistral Vibe - Architecture Overview
- Entry Points and Execution Flow
- Agent Conversation Loop
- Tool System Architecture
- Middleware Pipeline
- LLM Backend Abstraction
- Configuration System
- TUI Architecture and Event Flow
- Agent Profiles and Skills System
- Session Management
- Trusted Folders Security
- LLM Interaction Patterns, Guardrails & Techniques
- Mistral Vibe Architecture Documentation