AAS Core: Auditable skill selection and planning for Codex and Claude
A local control plane for Codex and Claude to select reusable skills, validate and preview plans without recommending or applying them for you.
GitHub sickn33/agentic-awesome-skills Updated 2026-10-21 Branch main Stars 48.1K Forks 7.0K
Python JavaScript Shell TypeScript HTML Agent skills Local MCP Codex Claude Workbench

🧭 Decision Guide

Why trending now: Cannot be determined from the provided material.

Try it if you

  • You use Codex or Claude and want to save a reviewable set of skills from 2,724+ SKILL.md playbooks.
    The README section “AAS Core: Agent-First Preview” says Codex or Claude can inspect the complete local catalog and record exact selections.
  • You need to check aas-stack.json and generate a plan preview before installation.
    The README section “AAS Core: Agent-First Preview” says aas stack validate checks the manifest and aas stack plan writes an immutable preview.
  • You use Antigravity and need selective activation because the skill context is overloaded.
    The README section “Bundles & Workflows” directs users to the selective activation guide.
  • You want to map roles or goals to bundles such as Web Wizard, Security Engineer, or OSS Maintainer.
    The README section “Bundles & Workflows” lists these bundle examples.

Skip it if you

  • You need Core to automatically rank or recommend candidate skills.
    The README explicitly says, “Core does not rank or recommend skills.”
  • You need supported-path Apply behavior or depend on automatic Recovery.
    The README says Apply and recovery are experimental and outside the supported preview.
  • Your manifest needs more than 128 skills.
    The README section “AAS Core: Agent-First Preview” sets a technical maximum of 128 skills for manifests.
  • You want structural validation alone to confirm semantic fit, compatibility, or operational safety.
    The README IMPORTANT notice says structural and identity validity does not establish semantic fit, compatibility, setup correctness, or operational safety.

Requirements

  • You need Codex or Claude and must configure the local MCP through the corresponding Codex setup or Claude Code setup.
  • The configuration command uses [email protected] and requires absolute paths for the Codex configuration file and cache-root.
  • Read the v19.12.1 AAS Core preview guide for approval, reconnection, validation, and planning details.
  • If installing known skills directly, you need a host skill directory such as --path .agents/skills in the command.

First step (verbatim from README)

npm exec --yes --ignore-scripts [email protected] -- aas mcp configure \
  --host codex \
  --scope user \
  --config /absolute/path/to/codex/config.toml \
  --cache-root /absolute/path/to/aas-cache

Watch out

  • The direct installer’s --dry-run preview does not consume or apply a Core plan.
    The README section “Install selected skills directly” explicitly says the direct installer does not consume or apply a Core plan.
  • Antigravity’s watched skill directory can overload when too many skills are installed.
    The README sections “Install selected skills directly” and “Bundles & Workflows” both describe Antigravity context overload.
  • Core preview validates manifests and structure but does not prove setup correctness.
    The README IMPORTANT warning explicitly excludes guarantees about setup correctness and operational safety.
  • The GitHub README tracks main, so Unreleased content may require a later release.
    The README says Unreleased features on main require a later release and that the versioned guide describes the published package.

Alternatives

  • agentic-awesome-skills direct installer：Use it when you already know the skill IDs and only need a --dry-run focused install without Core's selection and planning flow.
    Installation / Install selected skills directly
  • skills_index.json direct-host discovery：Use it when a host loads individual SKILL.md files and needs Stable Skills Manifest v1 instead of aas-stack.json.
    Stable Skills Manifest v1
  • Workbench：Use it when you only need to inspect saved Core stack and plan artifacts in browser memory without letting Workbench access the local filesystem.
    AAS Core: Agent-First Preview

Not stated in the README

  • The README does not specify minimum Node.js, npm, or Python versions.
  • The README does not provide a concrete compatibility matrix for Codex and Claude versions.
  • The README does not describe complete host compatibility for all 2,724+ SKILL.md files.
  • The README does not detail the exact limitations, failure behavior, or recovery process for experimental Apply and recovery.
  • The README does not state the per-skill semantic quality, maintenance frequency, or third-party content review standard for the 2,724+ skills.
  • The README does not explain permission boundaries or runtime isolation when skills are installed in production.

💡 Deep Analysis

6
No I want to build an internal integration on AAS Core’s MIT-licensed structure, but one project needs more than 128 skills. Can I use a single `aas-stack.json` for the entire collection?
For: An engineering integrator building an internal skill catalog from the MIT-licensed open structure while constrained by the 128-skill manifest limit

No. A single stack is not suitable for more than 128 skills because the project explicitly limits one manifest to 128 skills.

  • The project data lists the license as MIT, and the README exposes open files, indexes, and schemas, providing a basis for internal customization and integration.
  • However, the project insight states that a single manifest may contain at most 128 skills; larger collections must be split or redesigned.
  • AAS Core is intended for precise selection, stack validation, and plan preview, not for treating the entire 2,724+ catalog as one runtime collection.
  • For an internal catalog exceeding 128 skills, the integrator must define how to split stacks, coordinate multiple plans, and handle host context limits; the README does not provide a standard multi-stack execution model.
  • Project data: license is MIT License
  • Project insight, usage_limitations: a single manifest can technically contain at most 128 skills
  • README, “Browse 2,724+ Skills”: the complete library contains 2,724+ browsable skills
  • Project insight, solution_analysis: governance uses stack manifests, identity validation, structural validation, and immutable plans
Not stated in the README：The README does not say whether multiple stacks can be combined, validated, or represented by one approval digest；The README does not specify a recommended split strategy or host-side switching mechanism for collections over 128 skills
It depends I use skills in Antigravity, which watches the skills directory. With 2,724+ candidate SKILL.md files, can I use AAS Core to install only explicitly selected skills and avoid context overload?
For: An Antigravity user concerned that a watched skills directory will load too much context and who needs precise control over the installed set

It depends: AAS Core can substantially reduce the Antigravity installation set, but it cannot guarantee that selected skills will not cause context or behavioral interference.

  • The README explicitly warns that Antigravity’s watched skill directory can overload context and requires a selected set, a filter, or an explicit --all override.
  • The direct installer supports narrowing by skill ID, --risk, --category, and --tags, which helps avoid loading the full 2,724+ skill catalog.
  • The project also provides selective activation and overload-recovery guidance, showing that this host requires selective activation.
  • However, the project validates IDs, manifests, and structure; it does not prove semantic fit, runtime compatibility, or security for Antigravity.
  • README, “Installation”: Antigravity's watched skill directory can overload its context
  • README, “Bundles & Workflows”: use the installer's `--risk`, `--category`, and `--tags` filters
  • Project insight, usage_limitations: structural and identity validation does not prove semantic fit, cross-skill compatibility, or runtime safety
npm exec --yes --ignore-scripts [email protected] -- \
  agentic-awesome-skills --release 19.12.1 --path .agents/skills \
  --skills brainstorming,systematic-debugging --dry-run
Not stated in the README：The README does not quantify Antigravity’s context limit for different skill counts or SKILL.md sizes；The README does not say whether selective activation is automated by AAS Core or configured by the user in the host
Yes I use Codex and want local MCP to select skills from 2,724+ SKILL.md playbooks, but I must review the exact set before changing the target project. Is AAS Core suitable?
For: A Codex user who wants local MCP to select skills from project context while reviewing the exact skill set before any modification

Yes, because it is specifically designed as a local, agent-first selection and planning control plane for Codex.

  • The README states that AAS Core lets Codex or Claude search the complete local catalog and record the skills selected by the agent.
  • Its local MCP exposes catalog search and compose_stack, allowing selection of concrete skill IDs from project context instead of loading the entire catalog.
  • aas-stack.json, manifest validation, and aas stack plan preserve, validate, and preview the exact collection before changes are made.
  • However, Apply and recovery remain experimental, so it is suitable for selection, validation, and preview rather than a stable automatic execution loop.
  • README, “AAS Core — Agentic Awesome Skills”: AAS Core helps Codex or Claude search the complete local catalog, record the skills the agent chooses, and preview a plan
  • Project insight, solution_analysis: local MCP exposes catalog retrieval and compose_stack to Codex or Claude
  • README, “Current release”: Apply and recovery remain experimental
Not stated in the README：The README does not specify MCP performance, connection stability, or selection accuracy across Codex versions；The README does not explain how the agent resolves conflicting or semantically overlapping skills
Yes I manage a security-testing project and need to select from 2,724+ skills using `risk` and `category` filters, then inspect a dry run before writing to the host directory. Does AAS Core meet this installation-control requirement?
For: A security engineer managing 2,724+ skills who needs a minimal set filtered by risk and category with a dry-run before installation

Yes for discovery, scope reduction, and pre-installation preview, but no as a security-certification or automatic-approval system.

  • The README says the direct installer can narrow installation by skill ID, --risk, --category, or --tags, matching a need to filter by security attributes and category.
  • --dry-run previews a focused install before the flag is removed for the actual installation.
  • The project provides security guidance and distinguishes structural validity from semantic fit and runtime safety; this makes the installation scope visible but does not replace security review.
  • For natural-language skills involving command execution, file changes, credentials, or network access, the README does not promise runtime safety, so filtered results must not be treated as trusted security tools.
  • README, “Installation”: preview a focused install; `--dry-run`
  • README, “Bundles & Workflows”: use the installer's `--risk`, `--category`, and `--tags` filters
  • README, “Troubleshooting”: Security and antivirus alerts
  • Project insight, market_gap: structural validation is not semantic suitability or security certification
npm exec --yes --ignore-scripts [email protected] -- \
  agentic-awesome-skills --release 19.12.1 --path .agents/skills \
  --skills brainstorming,systematic-debugging --dry-run
Not stated in the README：The README does not define the `--risk` levels, defaults, or whether they are human-assigned security ratings；The README does not state whether each skill undergoes sandbox execution, malicious-command scanning, or dependency auditing
Yes I use Claude Code, need to save an exact skill set as `aas-stack.json`, review it with my team, and inspect the target path and change scope through a dry run before installation. Is this project suitable?
For: A Claude Code engineer who needs to commit a skill set to a team repository and review changes before installation

Yes, because the README provides a recordable, verifiable, and previewable skill-set workflow that matches team review requirements.

  • Claude Code is explicitly supported, with a dedicated Claude Code setup guide.
  • aas-stack.json stores an exact skill collection, while aas stack validate checks the manifest; this is more reviewable than committing only an installation command.
  • aas stack plan creates an immutable plan, and the Workbench can inspect saved stacks and plans in the browser without accessing the local filesystem.
  • The direct installer supports --dry-run, but the README explicitly says it does not consume or apply a Core plan. The team must therefore keep the Core planning path separate from direct installation.
  • README, “Support the Project”: follow Codex setup or Claude Code setup
  • Project insight, solution_analysis: aas-stack.json records selections and plan files plus approval digest support review
  • README, “Browse 2,724+ Skills”: Workbench inspects a saved Core stack and plan
  • README, “Installation”: The direct installer does not consume or apply a Core plan
Not stated in the README：The README does not specify how teams should version-lock stacks, resolve merge conflicts, or enforce approval permissions in Git；The README does not clarify whether the approval digest includes skill-content hashes or only a plan summary
No I maintain skills across Cursor, Gemini CLI, and Claude Code and want to reuse discovery logic through the machine-readable `skills_index.json`. Can AAS Core eliminate all compatibility differences among these hosts?
For: An integration developer moving skills among Cursor, Gemini CLI, and Claude Code who requires machine-readable manifests for each host

No, not if the goal is to eliminate every host-specific difference. It standardizes discovery and some governance, but it does not replace per-host validation of directories, configuration, or plugin compatibility.

  • The README provides skills_index.json, a v1 schema, and a compatibility mirror, which are suitable for machine-readable direct-host discovery integrations.
  • The project data and README list Cursor, Gemini CLI, and Claude Code among the supported compatibility targets, indicating broad coverage.
  • However, the README still provides separate host guides and documents plugin compatibility, configuration paths, and skill-directory issues, showing that host behavior is not fully abstracted.
  • The discovery manifest is also separate from aas-stack.json and the AAS Core catalog; copying one file cannot guarantee identical installation or loading behavior across hosts.
  • README, “Stable Skills Manifest v1”: skills_index.json, v1 schema, and compatibility mirror
  • Project data topics: cursor, gemini-cli, claude-code, and mcp
  • README, “Troubleshooting”: Plugin compatibility and installation; Windows context and truncation recovery
  • README, “Stable Skills Manifest v1”: This direct-host discovery manifest is separate from `aas-stack.json` and the AAS Core catalog
Not stated in the README：The README does not provide a complete configuration-path and version-compatibility matrix for Cursor, Gemini CLI, and Claude Code；The README does not specify whether the same SKILL.md has identical loading precedence and context behavior across all three hosts

✨ Highlights

  • The catalog contains 2,724+ installable SKILL.md playbooks.
  • compose_stack supports read-only validation and plan preview.
  • It supports Codex, Claude, local MCP, and Workbench.
  • Core explicitly does not rank or recommend skills.

🔧 Engineering

  • The local catalog searches complete skills and records exact selections.
  • aas stack validate checks manifests, while plan creates immutable previews.
  • Bundles group roles such as Web Wizard, while Workflows provide ordered playbooks.

⚠️ Risks

  • Apply and recovery remain experimental and are outside the supported preview path.
  • Structural and identity validity does not prove semantic fit or operational safety.
  • Too many installed skills can overload Antigravity's context.
  • Manifests have a technical maximum of 128 skills, and Core makes no recommendations.

👥 For who?

  • Development teams that need Codex or Claude to review skill selections.
  • Users managing local MCP, aas-stack.json, and plan previews.
  • Antigravity users who need selective skill activation.