Skip to main content

Use Agent Hub Claude-Compatible Skills

Use this path when you want to install an Agent Hub pack as a local Claude skill instead of fetching it at runtime over MCP.

Fastest Path​

Install Commands

Agent Hub (agent-hub v0.4.0)

This page reads optional skill and version query params and renders commands for the selected Claude skill bundle.

Package CLI

agent-hub install-skill agent-hub --version 0.4.0 --project
agent-hub install-skill agent-hub --version 0.4.0 --global

Repo Script

npm run install:claude-skill -- agent-hub --version 0.4.0 --project
npm run install:claude-skill -- agent-hub --version 0.4.0 --global

Manual bundle path: distributions/claude/agent-hub/0.4.0/

Install Via Prompt​

Install Via Prompt

Agent Hub (agent-hub v0.4.0)

Paste this into a local AI coding agent when you want the agent to choose the fastest available install path for this skill.

I want you to install the Agent Hub Claude skill for Agent Hub (agent-hub v0.4.0) in this local environment.

Use the fastest available path:
1. If the `agent-hub` CLI is available, run:
   agent-hub install-skill agent-hub --version 0.4.0 --project
2. Otherwise, if this Agent Hub repo is available locally, run:
   npm run install:claude-skill -- agent-hub --version 0.4.0 --project
3. Otherwise, download the generated bundle for agent-hub v0.4.0 and install it manually into:
   .claude/skills/agent-hub/
   Make sure `SKILL.md` is at the top level of that folder.

After installing:
- verify that .claude/skills/agent-hub/SKILL.md exists
- verify the references files are present
- tell me exactly what you changed
- if you could not complete the install, give me the exact manual steps for this environment

What This Installs​

Agent Hub installs a generated Claude skill bundle that looks like this:

distributions/claude/<tool>/<version>/
  SKILL.md
  references/
    overview.md
    api-groups.md
    workflows.md
    troubleshooting.md
  manifest.json

The installer copies that bundle into either:

.claude/skills/<tool>/

or:

~/.claude/skills/<tool>/

SKILL.md is the entrypoint. The references/ files hold the heavier material that would make the entrypoint too large or too noisy.

Choose Your Install Path​

Use the local installer​

Use the installer when you already have this repo or the Agent Hub package available locally and want the fastest Claude Code path.

Download the skill bundle​

Use the bundle download when you want to inspect the files first or copy them into another Claude-compatible environment manually.

When To Use MCP Instead​

Use MCP when you want runtime retrieval:

  • the agent should fetch packs only when a task needs them
  • you want version-aware retrieval without installing local files
  • you want one Agent Hub integration that supports many tools

Use a local Claude skill when:

  • the environment expects a local skill bundle
  • you want a file-based install
  • the local agent should load the skill from disk instead of calling MCP

These paths can coexist.

MCP is the just-in-time retrieval path. The generated Claude skill is the file-based install path.

Canonical Pack Vs Claude-Compatible Skill​

Canonical pack​

Use the canonical pack when you want:

  • the authoritative Agent Hub source
  • raw Markdown retrieved through agenthub_fetch
  • the artifact contributors edit and review directly

Claude-compatible skill​

Use the generated Claude-compatible skill when you want:

  • a local SKILL.md bundle
  • a file-based install path in a Claude-compatible environment
  • progressive disclosure through a short skill entrypoint plus supporting files

How To Fetch A Generated Skill Over MCP​

Agent Hub MCP now supports generated Claude-compatible skill retrieval.

Use:

  • agenthub_distributions to see which generated distributions exist
  • agenthub_fetch_distribution to fetch the full bundle
  • agenthub_fetch_distribution_file to fetch one file like SKILL.md

If you only want the canonical pack, continue using:

  • agenthub_fetch

Installing In Claude-Compatible Environments​

The exact install flow depends on the environment.

The important Agent Hub behavior is:

  • canonical packs remain the source of truth
  • generated skills are available in a Claude-compatible folder shape
  • MCP can also return the generated bundle programmatically

If your environment supports local skill installation, use the generated bundle. If it prefers runtime retrieval, use MCP instead.

Claude Code Local Install​

Claude Code discovers custom skills from the filesystem. In practice, that means you can install a generated Agent Hub skill into:

  • .claude/skills/ in the current project
  • ~/.claude/skills/ for a user-wide install

Claude Code Local Setup Check​

Agent Hub also includes a local preflight command that validates the filesystem-based pieces of Claude Code integration:

npm run check:claude-code-local-setup

This check:

  • installs the pilot skills into a temporary .claude/skills/ directory
  • adds the deployed Agent Hub MCP server to a temporary Claude Code project config
  • confirms Claude Code reports the MCP server as connected
  • cleans up the temporary project config afterward

If claude auth status is logged out, the check will still validate local install and MCP config, but it will skip the final runtime session probe. That is the remaining Milestone B blocker in issue #13.

Claude Code Runtime Validation​

Once Claude Code is logged in, Agent Hub also includes a runtime validation command that probes real skill usage with three pilot packs:

npm run check:claude-code-skill-runtime

By default it installs and probes:

  • react
  • playwright
  • supabase-js

The script:

  • creates a temporary Claude Code project
  • installs the selected generated skills
  • adds the deployed Agent Hub MCP server
  • runs three short Claude Code prompts against those installed skills
  • prints a markdown-ready report you can paste into issue #13

You can also target a subset explicitly:

npm run check:claude-code-skill-runtime -- react playwright

If you want the script to post the report directly to issue #13 after a successful logged-in run:

npm run check:claude-code-skill-runtime -- --issue 13

Contributor Rule: Edit Canonical, Regenerate Distribution​

When you improve a pack:

  1. edit agents/<tool>/<version>.md
  2. regenerate the Claude-compatible skill
  3. validate both artifacts

Example:

npm run validate:agent-pack -- agents/<tool>/0.4.0.md
npm run generate:claude-skill -- agents/<tool>/0.4.0.md
npm run validate:claude-skill -- distributions/claude/<tool>/0.4.0

If you are working on a pilot pack or changing the compiler, also run:

npm run check:claude-skill