Why AI Agent Skills Need a Universal Standard

The explosion of AI agent capabilities

A year ago, AI coding assistants could autocomplete functions and answer questions about your codebase. Today they can spin up databases, deploy infrastructure, send emails, browse the web, and orchestrate multi-step workflows across dozens of external services. Tools like Claude Code, Cursor, Codex CLI, and Windsurf have moved well beyond text generation into genuine agentic work.

Two protocols made this leap possible. MCP (Model Context Protocol) gives agents structured tool interfaces — JSON-RPC calls backed by JSON Schema, so an agent can call a Stripe API or query a Postgres database with typed inputs and predictable outputs. SKILL.md gives agents behavioral and procedural knowledge in plain markdown — how to write idiomatic React, when to prefer one library over another, or step-by-step instructions for debugging a memory leak.

Both approaches are powerful. Both are growing fast. And right now, they exist in completely separate worlds.

The fragmentation problem

MCP has become the de facto standard for giving AI agents structured tool access. GitHub, Stripe, Supabase, Cloudflare, Docker, Figma, Shopify — all have shipped official MCP servers. Community registries like Glama have indexed over 19,000 MCP servers. The ecosystem is thriving.

Meanwhile, SKILL.md has become the way developers encode procedural knowledge for agents. Skills.sh alone has indexed more than 83,000 SKILL.md files on GitHub. These range from coding style guides to full-blown debugging playbooks — things that can't be expressed as a tool call but dramatically improve what an agent can do.

Here's the problem: these two ecosystems don't talk to each other. Glama knows about MCP servers but ignores behavioral skills. Skills.sh knows about SKILL.md files but ignores MCP tools. If you build a weather tool with an MCP server and a SKILL.md that teaches the agent when to check weather versus use cached data, you have to publish to both registries, manage two discovery surfaces, and hope users find both halves. Most don't. The result is fragmented capabilities — agents that have the tool but not the knowledge to use it well, or the knowledge but not the tool to act on it.

What a universal standard looks like

We think the fix is straightforward: one manifest format that can describe any AI agent capability, whether it's an MCP tool, a behavioral skill, or both. We call it skill.json.

A skill.json is a superset manifest. It wraps MCP server configuration, references to SKILL.md files, agent compatibility metadata, and runtime information in a single, validated JSON document. Here's what a real one looks like:

{
  "$schema": "https://loaditout.ai/schema/skill/v1.json",
  "name": "weather-lookup",
  "version": "1.0.0",
  "description": "Get current weather for any location",
  "type": "hybrid",
  "compatibility": {
    "agents": ["claude-code", "cursor", "codex-cli"],
    "llm_providers": ["anthropic", "openai", "google"]
  },
  "mcp": {
    "transport": "stdio",
    "command": "node",
    "args": ["./dist/server.js"]
  },
  "skill_md": "./SKILL.md",
  "tags": ["weather", "api", "utility"]
}

The mcp field tells any MCP-compatible agent exactly how to start the server. The skill_md field points to behavioral instructions the agent should internalize. The compatibility block declares which agents and LLM providers this skill supports. One manifest, one source of truth, one CLI command to install: npx loaditout add user/weather-lookup.

Introducing Loaditout

Loaditout is the unified registry and runtime for AI agent capabilities. It bridges MCP tools, behavioral SKILL.md skills, and every major LLM provider into a single, searchable, installable ecosystem. Think of it as npm for AI agent skills — but format-agnostic and cross-platform by design.

We've already indexed over 400 skills from GitHub, including official MCP servers from Anthropic, GitHub, Stripe, Supabase, Cloudflare, Docker, AWS, Figma, and Shopify, alongside community favorites like Browser Use (55K+ stars), FastMCP, and the Playwright MCP server. Every skill is indexed with compatibility metadata so you can filter by agent (Claude Code, Cursor, Codex CLI) or provider (Anthropic, OpenAI, Google).

Loaditout is community-driven. You can submit your own skills by pasting a GitHub repo URL, upvote skills you find useful, and request skills you wish existed. The registry is open, the schema is open source, and there is no lock-in.

What's next

We're just getting started. On the roadmap: hosted skill execution so you can run MCP servers without installing anything locally, semantic search so you can find skills by describing what you need (“find me something that scrapes websites and returns clean markdown”), and an SDK for skill authors that makes publishing a skill as easy as npx loaditout publish.

If you build AI agents or use AI coding tools daily, we'd love your input. Join the waitlist on our homepage, submit your skills to the registry, or list your MCP server so others can discover it. The age of fragmented agent capabilities is ending. Let's build the universal standard together.