Skills Manager Server

A visual web interface for managing Claude MCP Skills with Claude Code CLI integration.

Project Overview

This project provides a browser-based UI for managing skills in an MCP (Model Context Protocol) server. Skills are markdown-based instruction sets that Claude can load on-demand to gain specialized knowledge.

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                        Skills Manager                           │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│   ┌──────────────────┐         ┌──────────────────────────┐    │
│   │  Claude Desktop  │         │     Browser UI           │    │
│   │  (MCP Client)    │         │  skills-manager.html     │    │
│   └────────┬─────────┘         └────────────┬─────────────┘    │
│            │                                │                   │
│            │ stdio/MCP                      │ HTTP REST         │
│            │                                │                   │
│   ┌────────▼─────────┐         ┌────────────▼─────────────┐    │
│   │   server.py      │         │  skills_manager_api.py   │    │
│   │   (MCP Server)   │         │  (Flask API :5050)       │    │
│   └────────┬─────────┘         └────────────┬─────────────┘    │
│            │                                │                   │
│            └────────────┬───────────────────┘                   │
│                         │                                       │
│                         ▼                                       │
│                  ┌──────────────┐                               │
│                  │   skills/    │                               │
│                  │  (folder)    │                               │
│                  └──────────────┘                               │
│                                                                 │
└─────────────────────────────────────────────────────────────────┘

File Structure

skills-mcp/
├── server.py                 # MCP server (FastMCP) - connects to Claude Desktop
├── skills_manager_api.py     # Flask HTTP API server (port 5050)
├── skills_manager_app.py     # Standalone all-in-one app (for .exe build)
├── skills-manager.html       # Web UI (single-file, Tailwind + Lucide)
├── requirements.txt          # Python dependencies
├── build.bat                 # Windows build script for .exe
├── skills/                   # Skills directory
│   ├── my-skill/
│   │   ├── SKILL.md          # Main skill content
│   │   ├── _meta.json        # Metadata (tags, description)
│   │   ├── scripts/          # Optional helper scripts
│   │   └── references/       # Optional reference docs
│   └── another-skill/
│       └── ...
└── dist/                     # Built executable + assets
    ├── SkillsManager.exe
    ├── skills-manager.html
    └── README.md

Setup

Prerequisites

• Python 3.10+
• Node.js (optional, for Claude Code CLI)
• Claude Code CLI (optional, for AI features)

Installation

cd skills-mcp
pip install -r requirements.txt

Running the API Server

python skills_manager_api.py

Opens at: http://localhost:5050

Running the MCP Server (for Claude Desktop)

The MCP server runs automatically when Claude Desktop starts (configured in claude_desktop_config.json).

Manual test:

python server.py

Daily Skills Catalogue Audit

The repo ships an automated review system that grades every skill daily and surfaces improvement opportunities. See reports/ for the latest output.

What it does

• Pass 1 (deterministic, every run): structural defects — missing files, description drift between _meta.json and SKILL.md frontmatter, broken references, unknown depends_on targets, oversized SKILL.md files, thin asset-less skills, duplicate sub-skill triggers, stale reviews.
• Pass 2 (one skill per run): content audit of the skill whose last_reviewed_at is oldest. Runs a deterministic rubric (description distinctness vs. tag-cluster siblings, library version mentions, asset richness, cross-references), proposes a relevance_tier (A/B/C/D), and optionally calls the Claude CLI for an LLM-assisted rubric.
• Pass 3 (weekly): catalogue-level review — sibling description overlap, router necessity, tag-cluster drift vs. last week's snapshot, source distribution.

Running locally

# Add the audit metadata fields once (idempotent):
python scripts/backfill_audit_meta.py

# Daily run (writes reports/YYYY-MM-DD.md):
python scripts/run_daily_audit.py

# With weekly Pass 3 + the LLM rubric:
python scripts/run_daily_audit.py --weekly --use-llm

The GitHub Actions workflow .github/workflows/skills-audit.yml runs the daily audit at 06:30 UTC and commits the resulting report. The Monday run also performs Pass 3.

Metadata fields added to _meta.json

Field	Type	Meaning
last_reviewed_at	ISO date or null	When Pass 2 last audited this skill
review_score	int 0–100 or null	Composite of distinctness + asset richness + freshness + cross-references
relevance_tier	"A" / "B" / "C" / "D" / null	A = irreplaceable, B = trim & keep, C = redundant with base-model output, D = obsolete

The Skills Manager UI shows these as a filter dropdown and a per-card badge.

Development

Key Files to Edit

File	Purpose
skills-manager.html	Frontend UI (vanilla JS, Tailwind CSS, Lucide icons)
skills_manager_api.py	Backend REST API (Flask)
server.py	MCP server for Claude Desktop integration
skills_manager_app.py	Standalone app (combines API + launcher for .exe)

API Endpoints

Method	Endpoint	Description
GET	/api/skills	List all skills
GET	/api/skills/<name>	Get skill details
POST	/api/skills	Create new skill
PUT	/api/skills/<name>	Update skill
DELETE	/api/skills/<name>	Delete skill
POST	/api/import/folder	Import skill from folder path
POST	/api/import/json	Import files via JSON
GET	/api/browse?path=	Browse filesystem
GET	/api/claude/status	Check Claude CLI availability
POST	/api/claude/run	Run Claude with prompt
POST	/api/claude/generate-skill	Generate skill with AI

Frontend Structure

The UI is a single HTML file with embedded CSS and JavaScript:

• Modals: Import, Folder Browser, File Upload, Generate, View, Claude Console
• State: skills[], selectedFolderPath, browsePathCache{}, etc.
• Key Functions:
  • browsePath(path) - Navigate filesystem
  • importSelectedFolder() - Copy folder to skills/
  • uploadFiles() - Upload files as new skill
  • generateSkill() - AI skill generation
  • runConsoleCommand() - Execute Claude prompts

Building the Executable

# Windows
build.bat

# Or manually
pip install pyinstaller
python -m PyInstaller --onefile --name "SkillsManager" --console ^
    --add-data "skills-manager.html;." skills_manager_app.py

Output: dist/SkillsManager.exe

Skill Format

SKILL.md Structure

---
name: skill-name
description: When Claude should use this skill
---

# Skill Name

## Overview
What this skill helps with.

## When to Use
- Trigger condition 1
- Trigger condition 2

## Quick Start
\`\`\`code
Example usage
\`\`\`

## Best Practices
- Practice 1
- Practice 2

## Examples
Practical examples here.

_meta.json Structure

{
  "name": "skill-name",
  "description": "One-line description",
  "tags": ["tag1", "tag2"],
  "sub_skills": [],
  "source": "imported"
}

Claude Desktop Configuration

Location: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "skills": {
      "command": "python",
      "args": ["C:/Users/<YOUR_USERNAME>/skills-mcp/server.py"]
    }
  }
}

MCP Tools (for Claude)

The MCP server exposes these tools to Claude:

Tool	Description
list_skills()	List all available skills
get_skill(name)	Load a skill's SKILL.md content
search_skills(query)	Search skills by metadata
search_content(query)	Full-text search across all skills
get_skills_batch([...])	Load multiple skills at once
reload_index()	Refresh skill index
validate_skills()	Check skill integrity

Known Issues / TODO

• Folder browser doesn't show drive letters on initial load (starts at root drives)
• No confirmation before overwriting existing skills
• File upload doesn't preserve folder structure from drag-drop
• Claude Console output doesn't syntax highlight
• No skill versioning or backup

Tech Stack

• Backend: Python 3.11, Flask, Flask-CORS, FastMCP
• Frontend: Vanilla JS, Tailwind CSS (CDN), Lucide Icons (CDN)
• Build: PyInstaller
• Protocol: MCP (Model Context Protocol)