AgentCore

The open standard for High-Density AI coding instructions. Make your AI smarter, faster, and more consistent.

Fork notice: AgentCore is a fork of the original agent-skills-standard, maintained by MacOsEsy at MacOsEsy/agentcore.

AgentCore is a modular framework to distribute, sync, and version-control engineering standards across all major AI agents (Cursor, Claude Code, GitHub Copilot, Windsurf, Kiro, Gemini, Antigravity, and custom LLM workflows), built on top of the original Agent Skills Standard.

💡 What is this?

Think of AgentCore as a universal instruction manual for your AI assistant.

Usually, when you use AI for work, you have to constantly remind it of your "house rules" (e.g., "Make sure to handle errors this way" or "Use this specific layout"). This project allows teams to package those rules into "Skills" that you can plug into any AI tool instantly.

Why does this matter?

• For Managers: Ensure your entire team writes high-quality code that follows your company's standards, regardless of which AI tool they use.
• For Non-IT Users: You don't need to know how the rules are written; you just run one command to "upgrade" your AI with professional-grade engineering knowledge.
• For Teams: No more "Hey, why did the AI write it this way?"—everyone's AI uses the same playbook.

⚡ The Problem: "The Context Wall"

Modern AI coding agents are powerful, but they struggle when managing project-wide rules:

1. Information Overload: Providing too many instructions at once confuses the AI and makes it forgetful.
2. Version Chaos: Every team member has a different version of "best practices" floating around their local computer.
3. Wasted Space: Long, wordy instructions "eat up" the AI's limited memory (the Context Window), making it more expensive and less effective.

🛠 The Solution: Digital DNA for AI

AgentCore treats instructions as versioned dependencies, much like software libraries.

• 🎯 Smart Loading: We use a "Search-on-Demand" pattern. The AI only looks at detailed examples when it specifically needs them, saving its memory for your actual code.
• 🚀 High-Density Language: We use a specialized "Compressed Syntax" that is 40% more efficient than normal English. This means the AI understands more while using fewer resources.
• 🔁 One-Click Sync: A single command ensures your AI tool stays up-to-date with your team's latest standards.

🛡️ Security & Trust

We understand that "injecting" instructions into your AI can sound risky. Here is how we keep you safe:

• No Code Execution: Skills are pure Markdown/JSON files. They contain text instructions for the AI, not executable code. They cannot run commands on your machine.
• Open Source: The entire registry is open source. You can audit every skill file on GitHub before using it.
• Sandboxed: The AgentCore CLI (agentcore) runs in user space to download text files. The "skills" themselves run inside the AI's isolated context window, not as OS processes.
• Privacy: We do not collect any code or project data. Feedback is only sent if you manually trigger the feedback command or strongly opt-in.

✨ Features

• 🛡️ Multi-Agent Support: Out-of-the-box mapping for Cursor, Claude Dev, GitHub Copilot, and more.
• 📦 Modular Registry: Don't load everything. Only enable the skills your project actually uses.
• ⚡ Proactive Activation (Universal): Generates a compressed index in AGENTS.md for 100% activation reliability across Cursor, Windsurf, Claude Code, and more.
• 🔄 Dynamic Re-detection: Automatically re-enables skills if matching dependencies are added.
• 🔒 Secure Overrides: Lock specific files so they never get overwritten.
• 📊 Semantic Tagging: Skills tagged with triggers for exact application.
• 🤖 Agent Workflows: Sync executable workflows (.md files) that agents can follow to perform complex tasks.

🛠️ CLI Commands

• init: Initialize a fresh .skillsrc config.
• sync: Sync skills from remote registry and automatically update the AGENTS.md index.
• validate: Validate skill format and token standards.
• feedback: Report skill violations or improvements.
• upgrade: Upgrade the CLI to the latest version.

✨ Token Economy & Optimization

To ensure AI efficiency, this project follows a strict Token Economy. Every skill is audited for its footprint in the AI's context window.

📏 Our Standards

• High-Density: Core rules in SKILL.md are kept under 70 lines.
• Efficiency: Target < 500 tokens per primary skill file.
• Progressive Disclosure: Heavy examples, checklists, and implementation guides are moved to the references/ directory and are only loaded by the agent when specific context matches.

🛠️ Token Calculation

We provide a built-in tool in the CLI to estimate and track the token footprint:

pnpm calculate-tokens

This command:

1. Scans all SKILL.md files in the registry.
2. Calculates character-based token estimates.
3. Updates skills/metadata.json with category metrics (total tokens, avg/skill, and identifying the largest skills).

By maintaining a "Lean Registry," we ensure that your AI assistant remains fast and focused, preserving the majority of its context window for your actual project code.

🔒 Privacy & Feedback

Feedback Reporter Skill

By default, the CLI syncs a common/feedback-reporter skill that enables you to report when AI makes mistakes or when skill guidance needs improvement. This helps us improve skills for everyone.

What Gets Shared (Only if You Report):

• Skill category and name
• Issue description (written by you or generated by AI)
• Skill Instruction: Exact quote from the skill that was violated
• Actual Action: What the AI did instead
• Decision Reason: Why the AI chose that approach
• Optional context (framework version, scenario)
• Optional AI Model name
• NO code, NO project details, NO personal information

How to Opt-Out: Add to .skillsrc:

skills:
  common:
    exclude: ['feedback-reporter']

Manual Feedback

If you notice a skill needs improvement, you can manually send feedback using (no installation required):

npx agentcore feedback

How it works:

• Cross-Platform: Works instantly on Windows, MacOS, and Linux via npx.
• The CLI attempts to submit your feedback automatically via our High-Density Feedback Backend (hosted on Render.com).
• No token needed: We handle the GitHub integration securely on the server.

Or via structured comments in your code:

// @agent-skills-feedback
// Skill: react/hooks
// Issue: AI suggested unsafe pattern
// Suggestion: Add guidance for this case

Privacy First: We never collect usage telemetry or analytics. Feedback is only shared if you explicitly trigger it.

🚀 Quick Start (Get running in 60s)

Consume engineering standards in your project instantly.

1. Run the CLI

npx agentcore@latest init

The interactive wizard will detect your stack and setup your .skillsrc.

2. Sync Standards

npx agentcore@latest sync

3. Automatic Activation

Whenever you run sync, the CLI automatically updates AGENTS.md in your project root with a compressed index of your enabled skills.

This AGENTS.md file serves as a universal entry point that helps all AI agents proactively understand when to trigger specific skills based on your project context.

⚙️ Configuration (.skillsrc)

The .skillsrc file allows you to customize how skills are synced to your project.

registry: https://github.com/MacOsEsy/agentcore
agents: [cursor, copilot, kiro]
skills:
  flutter:
    ref: flutter-v1.1.0
    # 🚫 Exclude specific sub-skills from being synced
    exclude: ['getx-navigation']
    # ➕ Include specific skills (supports cross-category 'category/skill' or 'category/*' syntax)
    include:
      - 'bloc-state-management'
      - 'react/hooks'
      - 'common/*'
    # 🔒 Protect local modifications from being overwritten
    custom_overrides: ['bloc-state-management']
  # 🤖 Optional: Sync workflows to .agent/workflows/
  workflows: true

Key Options

• exclude: A list of skill IDs to skip during synchronization.
• include: A list of skill IDs to fetch. Supports:
  • Relative Path: bloc-state-management (from current category)
  • Absolute Path: react/hooks (pull specific skill from another category)
  • Glob Path: common/* (pull ALL skills from another category)
• custom_overrides: A list of skill IDs that the CLI should never overwrite.
• ref: Specify a specific version or tag for the skills.

Project-Specific Skills

For project-specific rules that differ from or extend the standard registry skills, create local skills in your agent's skills directory (e.g., .github/skills/project/ for Copilot). These local skills will not be overwritten during sync operations.

Example:

Create .github/skills/project/your-project-standards/SKILL.md with your custom rules.

This approach ensures your project-specific standards are preserved while benefiting from upstream updates.

❓ Troubleshooting

Invalid .skillsrc

If the CLI complains about configuration format:

• Ensure registry is a valid URL.
• Ensure skills is a map of categories to config objects.
• Run npx agentcore@latest init to generate a fresh valid config.

🌍 Registry Ecosystem

AgentCore is designed to be the universal language for engineering standards.

🔹 Current Support (v1.6.3)

| Category | Key Modules | Version | Skills | Avg. Footprint | | :------------------------- | :---------------------------------------------------- | :------- | :----- | :------------- | | ☕ Spring Boot | Architecture, Security, Data, Test, Microservices | v1.0.0 | 10 | ~339 tokens | | 🌐 Common | SOLID, Security (SQLi/HTMLi), TDD, Anti-Patterns | v1.5.0 | 14 | ~494 tokens | | 🌐 Quality Engineering | Business Analysis, Jira, Zephyr, QA Automation | v1.0.0 | 4 | ~500 tokens | | 💙 Flutter | Clean Arch, BLoC, Riverpod, Testing, GetX, Nav v1 | v1.3.1 | 21 | ~422 tokens | | 🎯 Dart | Idiomatic Patterns, Advanced Tooling, Build Runner | v1.0.4 | 3 | ~435 tokens | | ☕ Java | Modern Syntax, Virtual Threads, Testing, Tooling | v1.0.0 | 5 | ~522 tokens | | 🔷 TypeScript | Type Safety, Security, Best Practices, Tooling | v1.0.6 | 4 | ~627 tokens | | 🟨 JavaScript | Modern ES2022+ Patterns, Async/Await, Functional | v1.0.0 | 5 | ~522 tokens | | ⚛️ React | Hooks, State Management, Performance, Security | v1.0.2 | 8 | ~396 tokens | | 📱 React Native | Arch, Perf, Navigation, Security, Deployment | v1.0.2 | 8 | ~396 tokens | | 🦁 NestJS | Architecture, Microservices, Security, CQRS, Scalling | v1.1.1 | 21 | ~476 tokens | | ▲ Next.js | App Router (v15), Caching (v16), Bundling, Debug | v1.1.1 | 17 | ~423 tokens | | 🐘 Laravel | Clean Arch, Eloquent, Security, Jobs, Redis, API | v1.0.0 | 10 | ~361 tokens | | 🐹 Golang | Clean Architecture, API Design, Concurrency, Security | v1.0.2 | 10 | ~351 tokens | | 🐘 Kotlin | Idiomatic Patterns, Coroutines, Flow, Tooling | v1.0.1 | 4 | ~453 tokens | | 🅰️ Angular | Standalone, Signals, Control Flow, SSR, Testing | v1.1.0 | 15 | ~276 tokens | | 🤖 Android | Architecture, Compose, DI, Perf, Testing, WorkManager | v1.0.2 | 22 | ~290 tokens | | 🍎 Swift | Language, Memory, Concurrency, SwiftUI, Testing | v1.0.0 | 8 | ~354 tokens | | 📱 iOS | Arch, UI, Lifecycle, Security, Perf, Networking, DI | v1.1.0 | 15 | ~404 tokens |

[!TIP] Quality Engineering (Opt-in): Advanced skills like business-analysis or zephyr-test-generation are in the quality-engineering category and are NOT synced by default to keep context lean. To use them, manually add the category to your .skillsrc:

skills:
  quality-engineering: { ref: quality-engineering-v1.0.0 }

[!TIP] Enterprise Ready: You can host your own Private Skills Registry on GitHub and point the CLI to it via the registry field in your .skillsrc.

📂 Standard Specification

The standard follows a strict directory structure designed for Token Economy.

skills/
└── flutter/                    # Category
    └── bloc-state-management/      # Skill
        ├── SKILL.md                # Core Rules (High Density)
        └── references/             # Heavy Examples (Loaded only on demand)

The CLI will sync this exact structure effectively to your agent configuration:

.cursor/skills/
└── flutter/
    └── bloc-state-management/

IDE Mapping

| Agent | Target Path | Integration Method | | :-------------- | :------------------ | :------------------------------------- | | Cursor | .cursor/skills/ | Automatic discovery via .cursorrules | | Trae | .trae/skills/ | Automatic discovery | | Claude Code | .claude/skills/ | Referenced in CLAUDE.md | | Copilot | .github/skills/ | Automatic discovery | | Kiro | .kiro/skills/ | Automatic discovery | | OpenAI | .codex/skills/ | Automatic discovery | | Antigravity | .agent/skills/ | Automatic discovery | | Gemini | .gemini/skills/ | Automatic discovery | | Roo Code | .roo/skills/ | Automatic discovery | | Windsurf | .windsurf/skills/ | Automatic discovery | | OpenCode | .opencode/skills/ | Automatic discovery |

🏗 Contributing & Development

Interested in adding standards for NestJS, Golang, or React? We follow a strict semantic versioning for every skill category.

1. Propose a Skill: Open an issue with your draft High-Density Content.
2. Develop Locally: Fork and add your category to skills/.
3. Submit PR: Our CI/CD will validate the metadata integrity before merging.

Local Development Testing

To test your changes locally before publishing to NPM:

# In your test project directory, point npx to your local CLI folder
npx /path/to/agent-skills-standard/cli sync

This is the recommended way to verify that skill injections and discovery bridges (like .cursorrules) work correctly in a real project environment.

📄 License & Credits

• License: MIT
• Original author: Hoang Nguyen
• Current maintainer: MacOsEsy