Support skill market search

[lironghai]

Current Problem / 当前问题
能够支持skill搜索

Proposed Solution / 建议方案
增加skill市场接入，侧边栏增加skill标签

Alternatives / 替代方案

Additional Context / 补充说明

[samanhappy]

可以说说你想达到什么样的使用效果吗，skills 基于文件夹，而 mcp 更多是为了连接其他服务，我想不出来要怎么结合

[josepcrespo]

Hi @samanhappy, although MCP servers and "skills" provide different functionalities (as you've already mentioned, MCPs allow AI agents to provide complex functionality with third-party tools, while "skills" are simple Markdown files that describe how an AI agent can perform certain tasks more accurately), I think it would be a good idea to add a "skills" marketplace to MCPHub (separate from the MCP functionalities that MCPHub currently provides).

How can we do this? As a first step, just as there's an integration of different MCP server marketplaces, why not integrate a "skills" marketplace where users can search for "skills"?

Other ways to expand MCPHub's integration/functionality for skills:

• Add semantic skills search to the marketplace (perhaps using cosine proximity supported by embeddings) and, if this is possible for the skills marketplace, why not adding this type of semantic search to the MCP server marketplaces as well?
• Add a button that makes it easy for users to install skills in their IDEs (VSCode and others), at least a "copy" button to copy the markdown of the skill.

If we consider this possibility, @lironghai, could you please tell us which APIs we could use to integrate a skills marketplace? From what source could we expose skills to MCPHub users? I know of a couple of GitHub repositories where "skills" are collected and published https://github.com/anthropics/skills/tree/main and https://github.com/github/awesome-copilot but perhaps you were thinking of using a specific source.

[samanhappy]

Hi @lironghai and @josepcrespo, Thank you for the detailed suggestions. I want to highlight a critical architectural challenge regarding the "installation" of Skills vs. MCP Servers.

MCPHub is primarily a self-hosted (Docker) solution.

• MCP Servers are easy to "install" because they run within the MCPHub container and are exposed via a unified endpoint.
• Skills, however, are Markdown files that must be physically present in the Agent's local environment (e.g., the local machine running Claude Code or Codex) to function.

If a user runs MCPHub in Docker, "installing" a skill into the container won't make it accessible to their local Agent. This creates a "last mile" delivery problem. Without solving this physical pathing issue, a simple "Install" button might be misleading for our Docker users. What do you think?

[josepcrespo]

Hi @samanhappy, thank you for raising this architectural concern — the "last mile delivery" problem is a real challenge. I've been thinking about how to address it, and I believe there are several complementary approaches that together could make a Skills marketplace viable in MCPHub without being misleading to Docker users.

---

Proposed Solution

Skills Source: SkillsMP API

I suggest using SkillsMP (they claim to have a database of almost 400 thousand skills) as the primary source for the marketplace. It provides a public REST API with both keyword search and AI semantic search endpoints:

GET https://skillsmp.com/api/v1/skills/search?q=<query>
GET https://skillsmp.com/api/v1/skills/ai-search?q=<query>

Each MCPHub user would provide their own SkillsMP API key (configurable via mcp_settings.json, environment variable, and a new "Skills" section on the /settings page in the MCPHub web panel), keeping us free from any centralized rate limit dependency (each API key has a limit of 500 requests per day). This is consistent with how MCPHub already handles other third-party credentials.

---

Solving the Last Mile Problem

I propose three complementary approaches that together cover most real-world usage scenarios:

---

1. Export Mechanisms (Immediate Value)

Rather than a misleading "Install" button, we offer clear, honest export actions once a user finds a skill they like:

• Copy Markdown — copies the raw skill content to clipboard, ready to paste anywhere.
• Download as .md — triggers a direct browser download of the skill file.
• Open via URL — a direct link to the skill's raw Markdown (/api/skills/:id/raw) that any tool or script can consume.

These are simple to implement, work correctly for both Docker and native users, and provide immediate utility with no risk of confusion.

---

2. Docker Volume Bridge (For Docker Users)

For users running MCPHub in Docker, we can document and support a volume mount pattern that bridges the container to the host filesystem. MCPHub would write downloaded skills into /app/skills-sync inside the container, which maps to the chosen path on the host — exactly where agents like Claude Code or Codex expect to find them. No manual copy needed.

User-wide installation (skills available to all projects on the machine):

# docker-compose.yml
services:
  mcphub:
    volumes:
      - ${HOME}/.skills:/app/skills-sync

Per-project installation (skills scoped to a specific project directory):

# docker-compose.yml
services:
  mcphub:
    volumes:
      - ${HOME}/Documents/GitHub/my-project/.github/skills:/app/skills-sync

The user can use any absolute path on the host machine. The target path inside the container (/app/skills-sync) stays the same in both cases — MCPHub doesn't need to know the distinction. The user simply maps it to whichever local directory makes sense for their workflow.

This approach is transparent and well-understood by Docker users, and it turns the Docker "barrier" into a first-class supported workflow.

---

3. Webhook / Sync Mechanism (Background Sync)

For users who want automation beyond manual export, we can build a lightweight opt-in sync system:

• MCPHub exposes a subscription endpoint where a local companion process (e.g., a small CLI tool or VSCode extension) can register itself as a subscriber.
• When a new skill is available or a user selects one to install, MCPHub sends a webhook notification to the registered local client.
• The local client receives the notification, downloads the skill from MCPHub's API, and saves it to the correct local path automatically.

The companion tool would support a configurable target path, allowing the user to choose between user-wide or per-project installation:

# User-wide: skill saved to ~/.skills/
mcphub-skills sync --target-path ~/.skills/

# Per-project: skill saved to specified absolute path
mcphub-skills sync --target-path ${HOME}/Documents/GitHub/my-project/.github/skills/

The user can specify any absolute path on their machine. MCPHub's internal path (/app/skills-sync) remains constant.

---

Summary

Approach	Best For	Effort
Copy / Download / Raw URL	All users, immediate needs	Low
Docker Volume Bridge	Docker users wanting automation (user-wide or per-project)	Low–Medium
Webhook / Sync Mechanism	Power users, automated workflows (user-wide or per-project)	Medium–High

None of these approaches would show a misleading "Install" button. Instead, the UI would adapt based on context, offering the most relevant option depending on how the user is running MCPHub.

---

Implementation Notes

A full Skills marketplace integration in MCPHub is a substantial feature encompassing API integration, UI components, database schema extensions, webhook infrastructure, and companion tooling. Given its scope, I propose marking this as a future roadmap item rather than prioritizing it immediately. This would allow us to:

• Complete other pending issues and features currently in development
• Gather more community feedback on Skills and their use cases (with IDEs like VSCode, CLI tools like Claude, others…)
• Ensure database and architecture needed for this feature implementation are well discussed and clear before starting to implement this huge new MCPHub feature

If we do decide to start on it, these three complementary approaches provide a clear path forward that respects the Docker constraint while delivering real value at each maturity stage.