bbeierle12-skill-mcp-claude

Skills Manager

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