AAS Core Turns Agent Skill Shopping Into a Reviewable Stack

Browsing a huge skill catalog usually ends the same way: the agent picks a few candidates mid-conversation, the picks scroll away, and nothing reproducible is left behind. Agentic Awesome Skills — 47k stars — builds its whole product around fixing that: the agent composes, you review, and the chosen set becomes a file you can validate, plan against, and reinstall.

Why This Skill Matters

The repository carries 2,134+ skills across development, testing, security, infrastructure, product, and marketing. On top of the catalog sits AAS Core, a local control plane with a clear division of labor: your agent (Codex or Claude) inspects the project, searches the complete local catalog, and chooses exact skill IDs. AAS never ranks or recommends. Its read-only MCP tools — including search_skills, get_skill, list_skill_files, read_skill_file, compose_stack, inspect_stack, and diff_stack — verify the agent-owned selection in memory, and a client or the aas CLI persists it as aas-stack.json with an immutable plan preview before any target change.

Two boundaries are worth knowing up front. Structural and identity validity does not certify semantic fit or operational safety — the README says so explicitly. And apply and recovery remain experimental, behind explicit opt-in flags and outside the supported safety claim. What you get today is search, selection, validation, and planning; execution stays a reviewed step.

Installation

For most hosts, the direct installer places skills where your agent watches:

npx agentic-awesome-skills --codex

--cursor, --gemini, and other per-host flags follow the same pattern. The default target is Antigravity, and there the installer refuses to clone or write unless you pass --skills, a metadata filter, or explicit --all — a full watched catalog can exhaust the agent's context or trigger a crash loop. Verify the result and run your first skill:

test -d ~/.agents/skills && echo "Skills installed in ~/.agents/skills"

Use @brainstorming to plan a SaaS MVP.

Real Workflow: Compose a Reviewed Stack for a Project

1. Configure the local AAS Core MCP for Codex or Claude, then describe the outcome you want. The agent searches the full catalog, compares candidates, and reads the selected instructions and support files.
2. Review the proposed aas-stack.json. The manifest preserves the exact agent-selected IDs — check prerequisites before approving.
3. Validate and preview without applying:

aas stack validate
aas stack plan

4. Preview the direct installer with those same IDs, then install for real:

npx agentic-awesome-skills --skills brainstorming,backend-dev-guidelines --dry-run
npx agentic-awesome-skills --skills brainstorming,backend-dev-guidelines

Real Workflow: Audit a Skill Before It Runs

Installation copies files; it does not execute their commands — but an agent may act on an installed skill later, so a review step earns its keep:

npx agentic-awesome-skills audit --skills brainstorming,backend-dev-guidelines

The audit reads the selected skill directories without executing them and reports command, network, credential, filesystem, privileged, destructive, symlink, and binary signals. The README is upfront that it is a review aid, not a safety certificate.

Tips

• A full-catalog install includes critical and authorized-use-only offensive instructions. Prefer an exact reviewed set over --all.
• If the job has a clear domain, start with a specialized plugin instead of the whole library — the README's plugin table lists 13 packs (AAS Web App Builder, AAS Security Engineer, AAS Data Analytics, and more), each bundling 8–10 skills.
• A stack manifest tops out at 128 skills — a technical maximum, not a recommendation.
• The hosted Skill Workbench imports and reviews stack and plan JSON in browser memory; it never touches your filesystem or installs anything.

When Not to Use This

If you already know the one skill you need, a single-skill direct install is simpler than composing a stack. And if you expect the catalog to guarantee that a skill fits your project, this is the wrong tool: AAS validates identity and structure, and the README leaves semantic suitability, compatibility, and setup correctness with you.

See the leaderboard for more skills.