Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.selftune.dev/llms.txt

Use this file to discover all available pages before exploring further.

​
What is selftune?

selftune is a local-first, self-improving skills toolkit for AI coding agents. It watches real agent sessions, detects missed skill triggers, grades execution quality, and evolves skill descriptions to match how users actually talk — automatically.

​
The problem

Agent skills are static but users are not. When you ship a skill, you write a description and hope it triggers when needed. There is no feedback loop. You don’t know whether skills are firing, under-firing, or missing entire categories of user intent. A user says “make me a slide deck” but the pptx skill stays silent because its description only matches “create a PowerPoint presentation.”

​
The solution

selftune closes this gap with a continuous feedback loop:
  1. Observe — Watch real agent sessions and log every user query
  2. Detect — Identify missed triggers and grade execution quality
  3. Evolve — Propose improved skill descriptions using real usage patterns
  4. Watch — Monitor for regressions and auto-rollback if needed

Get started in 2 minutes

Install selftune and see your first skill health report.

Build your first skill

Follow one end-to-end path from first draft to evolved description.

​
Key features

3-Tier Grading

Grade skills on trigger accuracy, process correctness, and output quality.

Autonomous Evolution

Pareto multi-candidate evolution with constitutional validation and auto-rollback.

Multi-Platform

Claude Code, Codex, OpenCode, OpenClaw, and Pi support.

Zero Dependencies

Pure TypeScript, MIT licensed, no API keys needed beyond your existing agent subscription.

Local Dashboard

Real-time SPA dashboard with skill health, evolution history, and SSE live updates.

Cloud Platform

Optional cloud dashboard for team-wide skill health monitoring and contributor signals.

​
How it works

User Query → Agent Session → selftune hooks capture data
                                    ↓
                              selftune sync normalizes
                                    ↓
                              selftune grade evaluates
                                    ↓
                            selftune evolve proposes improvements
                                    ↓
                            selftune watch monitors for regressions
                                    ↓
                              Auto-rollback if needed

Deep dive into the architecture

Understand the four layers: Capture, Normalize, Decide, Surface.