@netresearch/agent-skill-coordinator

Node Agent Skill Coordinator

A small Node package that discovers AI agent skills shipped inside installed npm packages and registers them in the project's AGENTS.md. Sibling of netresearch/composer-agent-skill-plugin — same skill format, same AGENTS.md block, just for the Node ecosystem.

What it does

When you npm install a package that declares one or more skills, this coordinator's postinstall hook scans node_modules, validates the skill metadata, and writes a managed block into AGENTS.md listing every discovered skill. Agents like Claude Code read that block to know which skills are available in the project.

Install

In a consuming project, add the coordinator and any skill packages as dev dependencies:

{
  "devDependencies": {
    "@netresearch/agent-skill-coordinator": "^0.1",
    "@you/some-skill-package": "^1.0"
  }
}

For pnpm, allowlist the postinstall script once:

{
  "pnpm": {
    "onlyBuiltDependencies": ["@netresearch/agent-skill-coordinator"]
  }
}

For Yarn Berry, set nodeLinker: node-modules in .yarnrc.yml (Plug'n'Play mode is not supported — there is no node_modules to scan).

After npm install completes, an AGENTS.md block listing all discovered skills is created/updated automatically.

Shipping a skill in your npm package

Skills are plain markdown files with YAML frontmatter (Anthropic skill convention).

SKILL.md:

---
name: my-skill
description: Short description of what this skill does and when to use it.
---

# My Skill

Detailed instructions, examples, and any bundled resources go here.

Frontmatter rules (mirror the Composer plugin):

• name (required): kebab-case, [a-z0-9-]{1,64}.
• description (required): ≤1024 characters, no control characters or bidi-override codepoints.

package.json declares the skill via aiAgentSkill (path, or array of paths) and includes the ai-agent-skill keyword for npm-search discoverability:

{
  "name": "@you/my-skill",
  "version": "1.0.0",
  "keywords": ["ai-agent-skill"],
  "aiAgentSkill": "SKILL.md",
  "files": ["SKILL.md"]
}

For multi-skill packages:

{
  "aiAgentSkill": ["skills/foo/SKILL.md", "skills/bar/SKILL.md"]
}

If aiAgentSkill is omitted, the coordinator falls back to SKILL.md at the package root (provided the ai-agent-skill keyword is present).

CLI

agent-skills list [--json]      List discovered skills
agent-skills read <name>        Print a skill's SKILL.md and metadata
agent-skills install            Regenerate AGENTS.md (manual trigger; same as postinstall)
agent-skills --help             Show help
agent-skills --version          Show version

The read command prints a header with absolute paths so an agent can cd to the base directory before invoking bundled scripts.

Programmatic API

import { discoverSkills, updateAgentsMd } from '@netresearch/agent-skill-coordinator';

const { skills } = await discoverSkills('./node_modules', {
  onWarning: (msg) => console.warn(msg),
});

await updateAgentsMd('./AGENTS.md', skills);

How the AGENTS.md block looks

The coordinator writes a <skills_system> block bracketed by <!-- SKILLS_TABLE_START --> / <!-- SKILLS_TABLE_END -->. Existing content outside the block is preserved. The XML structure is byte-identical to the Composer plugin's output (same wrapper element, same skill schema), so any agent that already reads the Composer plugin's output works here unchanged.

Compatibility

Package manager	Status
npm (≥7)	✅
pnpm	✅ (add coordinator to pnpm.onlyBuiltDependencies)
Yarn Classic	✅
Yarn Berry (nodeLinker: node-modules)	✅
Yarn Berry (default Plug'n'Play)	❌ (no node_modules to scan; use nodeLinker: node-modules)
Bun	✅

Hybrid PHP / JS projects

If your project also uses netresearch/composer-agent-skill-plugin, both tools maintain the same <skills_system> block in AGENTS.md. They overwrite each other on every run — last writer wins. For now, run them sequentially when you change skills on either side. Cross-ecosystem merging is a future enhancement.

Why a coordinator package and not postinstall per skill

• One install-time trust prompt covers all current and future skills.
• One AGENTS.md write per install, no race between concurrent skill packages.
• Skill packages stay pure data — no scripts, smaller security surface.

This mirrors how husky, simple-git-hooks, and other npm install-time tools work.

License

MIT — Copyright (c) 2026 Netresearch DTT GmbH. See LICENSE.