Content
# Skillhub MCP
<p align="center">
<img src="./assets/logo.png" alt="Skillhub MCP logo" width="160" />
</p>
[](https://pypi.org/project/skillhub-mcp/)
[](https://pypi.org/project/skillhub-mcp/)
## Links
- PyPI: https://pypi.org/project/skillhub-mcp/
- PyPI v1.0.1: https://pypi.org/project/skillhub-mcp/1.0.1/
- Skills directory: http://skills.214140846.net/
mcp-name: io.github.214140846/skillhub-mcp
You already have Claude-style skills (`SKILL.md`), but in practice you often hit a wall:
- your client speaks MCP, not Claude Skills
- your team uses multiple agents (Cursor, Copilot, Codex, etc.), so skills are painful to reuse across tools
- you want a more flexible way to organize and ship skills (nested folders, zip packaging)
**Skillhub MCP** bridges that gap: it turns Claude-style skills into MCP tools, so any MCP client can call the same skills.
> ⚠️ Experimental. Skills may contain scripts/resources. Treat them as untrusted and run with sandboxes/containers when possible.
## Is this an MCP server or an MCP client?
This project is an **MCP server**.
- **Skillhub MCP (this repo)**: runs as a server process and exposes tools/resources to clients.
- **MCP clients**: editors/agents like Cursor, Claude Code, Codex, etc. They start or connect to MCP servers.
## What You Get
- Cross-client reuse: install once, use from any MCP client
- Flexible packaging: nested directories, `.zip` and `.skill` archives
- Skill resources: expose scripts/datasets/examples as MCP resources (files the client can read)
- Resource fallback: a `fetch_resource` tool for clients without native MCP resource support
- Multiple transports: `stdio` (default), `http`, `sse`
## Quick Start
Default skills root: `~/.skillhub-mcp`
### uvx (recommended)
```json
{
"skillhub-mcp": {
"command": "uvx",
"args": ["skillhub-mcp@latest"]
}
}
```
Use a custom skills root:
```json
{
"skillhub-mcp": {
"command": "uvx",
"args": ["skillhub-mcp@latest", "/path/to/skills"]
}
}
```
## Install in Popular Editors (MCP Clients)
Below are minimal working examples for mainstream “vibe coding” editors.
### Cursor
Cursor supports configuring MCP servers via `mcp.json`. Add the following to your
global `~/.cursor/mcp.json` or project `.cursor/mcp.json`, then restart Cursor.
```json
{
"mcpServers": {
"skillhub-mcp": {
"type": "stdio",
"command": "uvx",
"args": ["skillhub-mcp@latest", "/path/to/skills"]
}
}
}
```
### Claude Code
Option A: configure via Claude Code CLI (recommended for quick setup):
```bash
claude mcp add --transport stdio skillhub-mcp -- uvx skillhub-mcp@latest /path/to/skills
```
Option B: project-scoped configuration via `.mcp.json` at your project root. You
may need to explicitly allow project MCP servers in `.claude/settings.json`.
`./.mcp.json`
```json
{
"mcpServers": {
"skillhub-mcp": {
"type": "stdio",
"command": "uvx",
"args": ["skillhub-mcp@latest", "/path/to/skills"]
}
}
}
```
`./.claude/settings.json` (approve only this server)
```json
{
"enabledMcpjsonServers": ["skillhub-mcp"]
}
```
### Codex (OpenAI)
Option A: use the Codex CLI to add a stdio MCP server:
```bash
codex mcp add skillhub-mcp -- uvx skillhub-mcp@latest /path/to/skills
```
Option B: edit `~/.codex/config.toml`:
```toml
[mcp_servers.skillhub-mcp]
command = "uvx"
args = ["skillhub-mcp@latest", "/path/to/skills"]
```
## Skill Format
Skillhub MCP discovers skills under the root directory (default `~/.skillhub-mcp`).
Each skill can be:
- a directory containing `SKILL.md`
- a `.zip` or `.skill` archive containing `SKILL.md` (at the archive root or
inside a single top-level folder)
All other files become downloadable MCP resources for your agent to read. Note:
Skillhub MCP does not execute scripts; the client decides whether/how to run them.
Example layout:
```text
~/.skillhub-mcp/
├── summarize-docs/
│ ├── SKILL.md
│ ├── summarize.py
│ └── prompts/example.txt
├── translate.zip
├── analyzer.skill
└── web-search/
└── SKILL.md
```
Archive rules:
```text
translate.zip
├── SKILL.md
└── helpers/
└── translate.js
```
```text
data-cleaner.zip
└── data-cleaner/
├── SKILL.md
└── clean.py
```
## Directory Structure: Skillhub MCP vs Claude Code
Claude Code expects a flat skills directory (each immediate subdirectory is one
skill). Skillhub MCP is more permissive:
- nested directories are discovered
- `.zip` / `.skill` packaged skills are supported
If you need Claude Code compatibility, keep the flat layout.
## CLI Reference
`skillhub-mcp [skills_root] [options]`
| Flag / Option | Description |
| --- | --- |
| positional `skills_root` | Optional skills directory (defaults to `~/.skillhub-mcp`). |
| `--transport {stdio,http,sse}` | Transport (default `stdio`). |
| `--host HOST` | Bind address for HTTP/SSE transports. |
| `--port PORT` | Port for HTTP/SSE transports. |
| `--path PATH` | URL path for HTTP transport. |
| `--list-skills` | List discovered skills and exit. |
| `--verbose` | Emit debug logging. |
| `--log` | Mirror verbose logs to `/tmp/skillhub-mcp.log`. |
## Safety Notes
- Skills are not "just prompts": they can include scripts and arbitrary files.
- Skillhub MCP does not run scripts, but your client might. Prefer running in a sandbox/container.
## Language
- English: `README.md`
- 中文: `README.zh-CN.md`
## About the Author
I focus on **AI SaaS going global**, covering the full journey from **idea validation and vibe coding** to **product development, infrastructure, SEO, backlinks, and growth experiments**.
Everything shared here comes from real projects, real traffic, and real revenue attempts.
- **Feishu Knowledge Base**:
[Thor’s AI Going-Global Content Planning](https://my.feishu.cn/wiki/space/7271588985498140676?ccm_open_type=lark_wiki_spaceLink&open_tab_from=wiki_home)
A structured knowledge base documenting hands-on experience in AI product overseas expansion, including demand discovery, execution strategies, and common pitfalls.
- **Blog**:
[Thor-AI Blog](https://www.notion.so/Thor-AI-2eaf0388ab4680d0a98bedc8d290e1be?pvs=21)
Long-form notes and case studies on building, launching, and iterating AI products in public.
- **Open-source Project (High Star)**:
**Smart Campus System**
- GitHub: https://github.com/214140846/TOGO_School_Miniprograme
- Gitee: https://gitee.com/zengyunengineer/TOGO_School_Miniprograme
- **Social**:
[Jike](https://web.okjike.com/u/159D450D-2193-4739-8825-AA8EBEC2E9B4)
Sharing real-time thoughts on indie hacking, AI tools, and product growth.
- **Product**:
- **AI Video Generation Platform**:
[Sora 2](https://sora2.cloud/)
[Sora 2 ai](https://sora2.cloud/home)
An online platform for AI-powered video generation, focused on practical use cases and real user workflows.
- **AI Video & Image Generation**:
[AI Video & Image Collection](https://ricebowl.ai/)
Model pages:
- [Grok Video](https://ricebowl.ai/m/grok-video)
- [Sora 2](https://ricebowl.ai/m/sora/sora-2)
- [Veo 3.1](https://ricebowl.ai/m/veo/veo-3-1)
- [Veo 3](https://ricebowl.ai/m/veo/veo-3)
- [Veo 2](https://ricebowl.ai/m/veo/veo-2)
- [Kling 2.6](https://ricebowl.ai/m/kling-2-6)
- [Wan 2.5](https://ricebowl.ai/m/wan/wan-2-5)
- [Seedance](https://ricebowl.ai/m/seedance)
- [Nano Banana 2](https://ricebowl.ai/m/nano-banana-2)
- [Nano Banana Pro](https://ricebowl.ai/m/nano-banana-pro)
A curated collection of AI video and image generation tools, experiments, and capability tracking.
- **AI Video & Image Collection**:
[https://www.notion.so/2e7600937cb3808c818efe79141f7ee6](https://www.notion.so/2e7600937cb3808c818efe79141f7ee6?pvs=21)
