AI Skill Package Managers
The "npm for AI Skills"

By raw popularity, superpowers is the overwhelming category king with 266K+ stars. It's a full agentic skill framework that provides a DSL for defining AI agent capabilities, skill composition, and multi-agent orchestration. While not a "package manager" per se, superpowers includes skill packaging, dependency resolution, and discovery mechanisms.

Superpowers is most useful as a skill development framework — you build skills using its DSL and deploy them to agents. It lacks the registry/publishing workflow that defines npm, but its massive community and mature tooling make it the de facto standard for building agentic skills today.

reverse-skill takes a unique approach: it reverse-engineers how AI agents discover and invoke skills, then provides a routing layer that maps skill requests to the right handler. Think of it as a skill-oriented reverse proxy for AI agents.

It's particularly useful for teams running multiple agents (Claude + Codex + Gemini) who need a unified skill dispatch layer. However, it's a router, not a package manager — it doesn't help you publish, version, or share skills.

skill-recorder is Microsoft's tool for recording on-screen workflows and publishing them as CLI skills usable via Copilot. It captures GUI interactions (clicks, typing, navigation) and converts them into a reusable skill that can be triggered by natural language — essentially macro recording for AI agents.

While innovative, skill-recorder is tightly coupled to the Microsoft/Copilot ecosystem. Skills created with it are not portable to Claude Code, Cursor, or Codex CLI. It's also more of a skill authoring tool than a package manager — there's no registry, no versioning, and no dependency management.

Codex as MCP bridges Codex CLI with the Model Context Protocol (MCP) ecosystem. It runs Codex CLI as an MCP server, allowing other MCP-compatible agents (Claude, Cursor, etc.) to use Codex CLI's capabilities and vice versa. This is a protocol-level bridge rather than a skill package manager.

It's an important piece of the puzzle — MCP is the emerging standard for tool interoperability, and any serious "npm for AI skills" will need to be MCP-native. Codex as MCP demonstrates the bridging pattern but doesn't address packaging, publishing, or discovery of skills.

memmy-agent is a personal AI agent platform that runs locally with a persistent memory hub. It manages skills as part of its memory architecture — skills are stored, recalled, and composed based on the agent's memory of past interactions. This is a memory-first approach to skill management.

Memmy's skill management is deeply tied to its memory system. Skills are not standalone packages but are embedded in the agent's knowledge graph. This makes it powerful for personal use but less suitable for publishing or sharing skills across teams.

agents-council is a multi-agent collaboration framework that includes skill management as a core feature. It enables multiple AI agents to coordinate, share skills, and delegate tasks. Think of it as a skill governance layer for agent teams.

Its skill management is focused on coordination — who has which skill, how skills are routed between agents, and how skill conflicts are resolved. It's less about the npm-style publishing workflow and more about runtime skill orchestration.

agent-manager provides a terminal UI (TUI) for managing AI agent sessions and their associated skills. It's a dashboard for your agents — you can see which skills are loaded, enable/disable them, and monitor agent-skill interactions in real time.

While useful for operational management, agent-manager is a control panel, not a package manager. It assumes skills already exist and focuses on runtime management rather than the publish/install/discover workflow that defines npm.

Claude Code CLI includes a built-in skill system based on CLAUDE.md files, custom slash commands, and skill definitions in the project's .claude/ directory. Skills are defined as markdown files with structured frontmatter, and Claude Code discovers them automatically.

The system is elegant but isolated — skills are tied to Claude Code's internal agent architecture. There's no registry, no versioning, and no way to publish or share skills beyond copying files. It's a skill system, not a skill package manager.

Codex CLI has its own built-in skill system using rules files, prompt templates, and configuration in .codex/ directories. Skills are defined as structured prompt files that Codex CLI uses to guide its behavior across sessions.

Like Claude Code's system, Codex CLI skills are platform-specific. They work well within the Codex ecosystem but cannot be shared with or used by other agent platforms. The lack of a standardized format means the two built-in systems are effectively incompatible with each other — despite both solving the same problem.