AAS Core / Agentic Awesome Skills

Quick Verdict

AAS Core / Agentic Awesome Skills is for developers who use several coding agents and need a reproducible way to discover, select, validate, and review skill stacks. The canonical project has moved from the old Antigravity Awesome Skills name to the agentic-awesome-skills repository and npm package. Release 15.1.0 describes a catalog of more than 1,969 skills plus a local MCP server, the aas CLI, stack manifests, selection evidence, immutable plan previews, specialized plugins, and a browser-based Workbench.

Its most important boundary is easy to miss: AAS Core does not recommend or certify the best skills. Claude, Codex, or another host inspects the project and selects exact IDs; Core verifies catalog identity and structure. A structurally valid stack can still be redundant, incompatible, unsafe, or badly suited to the work. Search, inspection, composition, validation, planning, and diagnosis are the supported preview path. Apply and recovery remain experimental. Use AAS as a reviewable control plane, not permission for an agent to install thousands of instructions automatically.

Best For

• Multi-agent developers reusing SKILL.md workflows across Claude Code, Codex, Cursor, Gemini CLI, OpenCode, and related hosts.
• Engineering leads who want a durable record of selected skills, declared coverage, evidence, and planned target changes.
• Security-conscious teams willing to separate read-only discovery from approved writes and runtime execution.
• Skill authors and integrators comparing metadata, distribution formats, plugins, bundles, and workflows.
• Not ideal for users who need only one known skill, cannot inspect third-party instructions, or expect automatic semantic and security certification.

Key Features

• Complete local catalog: agents can search and read more than 1,969 skills without an AAS service inspecting the source repository.
• Read-only local MCP: tools include skill search, retrieval, stack composition, inspection, diffs, and selection-evidence export or inspection.
• Reproducible manifests: exact agent-selected IDs can be stored in aas-stack.json while process evidence remains in a separate sidecar.
• Validation and immutable planning: the CLI checks a proposal and creates a target-specific preview before materializing payloads.
• Multiple delivery surfaces: direct installs, focused plugins, bundles, workflows, and the full catalog address different context budgets.
• Workbench review: the hosted interface imports stack and plan artifacts in browser memory; it is a review surface, not a hosted control plane.
• Model-aware boundary: the host model chooses skills. Core does not promise that Claude and Codex will understand a project or select candidates identically.

Use Cases

• Ask an agent to inspect architecture, domain behavior, tests, security, deployment, and maintenance, then search multiple candidates for each relevant capability.
• Persist a reviewed skill set as a versioned manifest instead of relying on an undocumented conversation or a mutable directory copy.
• Preview the install, update, and removal plan for Claude Code, Codex, Cursor, Gemini CLI, OpenCode, or a custom path.
• Start with a focused security, web, data, QA, DevOps, or agent/MCP plugin rather than loading the complete catalog into a context-sensitive host.
• Audit duplicated skills, source attribution, risk metadata, external scripts, and host assumptions before adoption.

Pricing

The repository and npm package are free. Original code and tooling use the MIT license. Original documentation and other non-code written content use CC BY 4.0 unless a more specific notice applies. Third-party skills keep their upstream licenses and exceptions, so neither the repository’s MIT file nor its content license grants blanket permission to redistribute everything. Commercial teams should preserve attribution and review the source ledger for each selected payload.

AAS itself does not charge for model usage. Project inspection, candidate comparison, and workflow execution consume the subscription or API quota of the host model. npm and GitHub downloads also have network availability and rate limits, and the full package is large enough to create storage, scanning, update, and context costs. Pin a 15.x release, prefer its release-pinned installer behavior, and use dry-run or plan output instead of following main without review. The older antigravity-awesome-skills package remains on a 13.x compatibility line and is no longer the canonical Core package.

Pros

• Turns ad hoc skill copying into manifests, evidence, validation, diffs, and reviewable plans.
• Local discovery keeps the project-data boundary clearer than a hosted repository scanner.
• Broad host support and several packaging forms fit both narrow and expansive workflows.
• Focused plugins reduce context pressure when the full catalog would be counterproductive.
• Documentation explicitly separates identity validation from semantic fit and operational safety.

Cons

• Nearly two thousand skills create substantial selection, duplication, context, and supply-chain review work.
• Core is still an agent-first preview, while apply and recovery are outside its stable safety claim.
• The local stdio server is not a multi-user authentication or authorization system.
• Host support differs in paths, triggers, commands, permissions, and model behavior.
• MIT and CC BY cover original project material only; third-party licensing remains a per-skill obligation.

The core runs locally, but installation and updates require npm and GitHub connectivity. It does not require an AAS account or key; authentication, model access, and billing belong to the coding-agent host. That convenience should not erase the threat model. A third-party SKILL.md is executable influence over an agent: treat instructions, scripts, links, upload requests, and shell commands as untrusted supply-chain input. Inspect diffs and use least privilege before letting any selected skill touch credentials, code, or deployment systems.

Alternatives

FAQ

Is AAS Core free?

Yes. The project is free, while Claude, Codex, or other model usage may consume paid subscriptions or API credits.

Should I still install Antigravity Awesome Skills?

New deployments should use the canonical agentic-awesome-skills package. The old name identifies an earlier compatibility line and does not represent current 15.x Core.

Does AAS upload my repository?

The documented Core flow has the local host agent inspect the project and query a local catalog MCP. Review the data policy of the host model separately.

Does the local MCP require an account or API key?

No AAS account is required for local stdio. npm, GitHub, and the coding-agent host may have their own authentication and quotas.

Does a valid stack mean every skill is safe?

No. Validation confirms identities and structure, not semantic suitability, setup correctness, licensing, script safety, or permission boundaries.

Should I install the complete catalog?

Only when broad local discovery is worth the cost. Most projects are easier to audit with a focused plugin or a small, explicitly selected stack.

Bottom Line

AAS Core is most useful as governance for agent-owned selection: local search, exact IDs, persistent evidence, structural validation, and a plan people can review. It is not a ranking engine, security scanner, license clearance service, or substitute for understanding the host model. Pin a Core-capable release, keep discovery read-only, inspect every third-party instruction, prefer small stacks, and require approval before writes. Used that way, it makes skill adoption more reproducible without pretending that catalog membership creates trust.