skillhub

Trusted skill marketplace for AI agents. One manifest, many runtimes.

A skill is a small tool that an AI agent (Hermes, Claude Code, Codex, Cursor) can call as a black box. skillhub is the place where:

• maintainers publish skills once, in a universal skill.yaml,

• agents and humans discover them via skillhub search,

• they install into the right runtime with skillhub install <name> <runtime>.

No domain. No accounts. No money involved — yet. Just the protocol, the CLI, and a registry of 20 seed skills.

Status

v0.0.1 — local CLI, no network, no payments. The goal of v0.0.1 is to prove the manifest format and the discoverability workflow before we add infrastructure.

Related MCP server: aiskillstore

What's here

Install (local)

git clone <repo>
cd skillhub
python -m venv .venv && source .venv/bin/activate
pip install -e .
skillhub search "pdf"

Usage

# Search the local registry
skillhub search "search"             # human table
skillhub search "pdf" --json         # agent-friendly, one JSON per line

# Show one skill in detail
skillhub show pdf-md

# Install into a runtime
skillhub install pdf-md --runtime hermes
# or
skillhub install pdf-md -r claude-code

# Validate your own skill.yaml
skillhub validate ./my-skill/skill.yaml

As an MCP server (skillhub-mcp)

For AI agents that speak MCP (Claude Code, Hermes, Codex, Cursor), skillhub ships its own marketplace as a server. Connect once, then search/show/install/validate skills as tool calls — no copy-paste, no scraping.

{
  "mcpServers": {
    "skillhub": {
      "command": "python",
      "args": ["-m", "skillhub_mcp.server"],
      "cwd": "/path/to/skillhub"
    }
  }
}

The server exposes 11 tools — full agent lifecycle:

The retention loop is built in: every install writes to a local profile; every successful run is recorded as a rate; the next stats call surfaces those signals back. So agents get more confident about skills over time — not less.

See src/skillhub_mcp/server.py for the full schema.

• Maintainers write skill.yaml once; the CLI compiles to runtime layouts.

• Agents find skills via skillhub search --json instead of web scraping.

• Humans get a trust_score per skill — derived from real signals, not stars.

• Everyone agrees on the same skill.yaml schema, so we don't fork five copies of the same SKILL.md across runtimes.

Roadmap (no dates)

• v0.1: schema stable, security scanner v1, real registry updater.

• v0.2: trust score from install_success_rate (live telemetry).

• v0.3: in-agent MCP-server (skillhub-mcp) so any MCP-capable agent can discover/install skills via tool calls.

• v1.0: featured/verified tiers, paid placements, the actual marketplace.

License

MIT. See headers in source files.