The master observes before teaching.

AI coding tools are getting louder. More suggestions, more autocompletes, more interrupting. Sensei moves the other way.

Sensei is a patient observer. It sits beside your editor and AI assistants, watching the shape of each session — the prompts, the responses, the corrections. It speaks rarely, and only when it has something specific to say. Most days it is completely silent — and that is the feature.

The kanji throughout the app are not decoration. Each one names a phase of practice: 観 observation, 察 recognition, 覚 adoption, 練 refinement. They are what we ask of the user, and what we ask of ourselves.

Sensei exists to close the feedback loop between developers and their AI tools. Today, teams adopt AI assistants and hope for the best. There's no visibility into what's working, what's not, and where the friction lives. Sensei provides that visibility.

It tracks First-Try Rate — how often a session completes without corrections — per project, per module, per pattern. It detects recurring idioms and anti-patterns. It surfaces coaching recommendations with projected impact. And it does all of this locally, without sending a single byte off your machine.

Sensei is four components: a Rust daemon (senseid) that indexes your codebase and serves an HTTP API; an MCP server (sensei-mcp) that translates AI tool calls into daemon requests; a CLI (sensei) for setup, scanning, and management; and a Tauri desktop app for visual observation and configuration.

Data lives in a local PostgreSQL database with pgvector for semantic search embeddings. Ollama provides optional local inference for pattern detection, code similarity, prompt classification, and embedding generation — recommended models are chosen based on your hardware. All features degrade gracefully without Ollama.

The system supports multiple AI platforms simultaneously through a capability registry that adapts features to each platform's strengths.

Sensei uses Ollama for tasks that don't need cloud APIs: pattern detection, code similarity analysis, semantic search embeddings, prompt classification, and docstring generation. This keeps routine analysis local and reduces API costs.

Hardware-aware model selection recommends the right models for your machine:

When multiple models are available, sensei runs a consensus panel — models debate insights for root cause analysis and impact prediction. The reasoning is transparent and visible in the desktop app.