AAS Core — Agentic Awesome Skills

A complete local skill catalog for coding agents—from project inspection and agent-owned selection to a reproducible, reviewable plan.

Current release: V15.1.0. This release includes AAS Core for complete local catalog search, agent-owned selection, manifest validation, planning, and diagnosis. Apply and recovery remain experimental and outside the supported preview path.

Codex or Claude inspects your project, enumerates its primary capabilities, searches and compares candidates across the complete local AAS catalog, and chooses the exact skills. Core imposes no semantic policy that favors a small stack; the manifest format has an explicit technical maximum of 128 skills. All 1,968 skills in the current catalog remain individually searchable, readable, and selectable. AAS Core does not rank or recommend skills. Its read-only compose_stack tool validates and returns the agent-owned manifest in memory; a client or the aas CLI persists the reviewed stack and its optional selection-evidence sidecar.

Read the AAS Core preview guide →

Project
  -> inspected by Codex or Claude (not by AAS)
  -> agent searches and reads the complete local catalog
  -> AAS MCP (local stdio, read-only)
  -> Codex or Claude chooses exact skill IDs
  -> compose_stack validates the selection in memory (read-only)
  -> client or AAS CLI persists aas-stack.json and optional evidence
  -> AAS CLI validate + immutable plan preview
  -> human review (optionally in Workbench)

The 1,967+ reusable SKILL.md playbooks, specialized plugins, bundles, workflows, and direct installers remain important. They are the content, curation, distribution, and compatibility layers around AAS Core—not competing primary products.

This is an independent community project. It is not affiliated with, sponsored by, endorsed by, or authorized by Google. Google, Antigravity, Gemini, and related product names are referenced only to describe compatibility and install targets. The GitHub repository is canonical; the hosted catalog and browser-local Workbench are companion discovery and review surfaces, not a hosted control plane.