The design
The design
Agent Skills DB is an implementation of the design laid out in Maximilian Bladt's paper on shared skill libraries for coding agents, built on the Agent Skills open standard for the skill files themselves.
The paper's central argument: a team's agent knowledge should live in a database with operations, not a repository of files. A repository can store skill files, but the operations a living library needs — merge two overlapping skills in one transaction, adjust who may see what, age out what nobody uses, answer "which version was active in March" — are database operations. Git gives you the file history; it does not give you the library.
From the paper, this implementation keeps:
- Skills as versioned rows, bodies deduplicated by hash — every change
- Distribution as sync, not clone — each machine holds exactly what its
- Per-developer review — a new skill activates per machine, when its
- Merge as a first-class transaction — umbrella published, sources
- Deterministic checks before publication — code, not model judgment,
- Usage telemetry driving lifecycle — skills nobody loads go stale at
a new immutable version, rollback to any of them
operator is entitled to, placed into each coding agent's own folder
operator says so; the library converges without central approval
archived, redirects recorded: all or none
decides what may leave a machine
thirty days and archive at ninety, recoverably; locked standards are exempt
Where this implementation deliberately deviates, the reasons are recorded in the repository's architecture decision records.