luongnv89/asm

What it solves

agent-skill-manager (asm) solves the problem of fragmented skill management across multiple AI agent tools. Instead of manually copying skill files into various hidden directories for tools like Claude Code, Cursor, and Codex, it provides a centralized CLI to install, search, audit, and organize skills across 21 different agent providers.

It specifically addresses the "attention budget" problem where every installed skill's description consumes tokens in the agent's system prompt on every message. asm allows users to track these costs and demote skills to lower residency tiers to save context window space.

How it works

asm acts as a management layer between a skill registry/GitHub and the specific directory structures required by various AI agent providers.

1. Installation & Management: It validates SKILL.md files, scans for security risks (like shell execution or credential exposure), and installs skills into the correct provider paths.
2. Residency Tiers: It manages how skills are stored to control their prompt cost:
   • Installed: Auto-triggers and always resident in context.
   • Saved: Stored in a local library but not active until activated.
   • Disabled: Kept on disk but not resident.
   • Reference: Fetched on-demand via asm get without being permanently installed.
3. Agent Integration: The CLI is designed for AI agents themselves to use, providing --json and --yes flags for non-interactive, machine-readable output.

Who it’s for

• AI Agent Users: People using multiple agentic tools (e.g., Claude Code, Cursor, Aider) who want a single way to manage their capabilities.
• Agent Developers: Those building AI agents that need a way to dynamically discover and acquire skills or dependencies just-in-time.
• Power Users: Users concerned with optimizing their LLM context window (attention budget) by managing resident skill descriptions.

Highlights

• Cross-Provider Support: Manages skills for 21 different AI agent tools from one interface.
• Context Optimization: Tools to calculate token costs for resident descriptions and audit which skills to demote to save tokens.
• Security Scanning: Pre-install scans for network access, shell execution, and credential exposure.
• Agent-Ready CLI: Full JSON output support for seamless integration into agent tool-calling workflows.
• Just-in-Time Delivery: Ability to fetch a skill's body via asm get without installing it, treating the tool as delivery infrastructure.
• Dependency Management: Supports optional skill dependencies that can be acquired and released on a per-session basis.