MCP Radar 🛰️

MCP Server quality radar — find the trustworthy MCP server out of 20,000+, and avoid the 98% that nobody ever installs.

Dataset: 19,947 approved MCP servers (scraped from mcp-marketplace.io, 2026-08-25) The problem: 98% of servers have zero installs, 91% pile into "Developer Tools", only 0.1% are paid, and a single server's schema can burn 47K tokens before you even ask a question.

What it does

Works inside any MCP client (Claude Desktop / Cursor / Claude Code / etc.):

1. search_servers — search high-quality servers by keyword/category, ranked by an open quality score

2. audit_server — audit any single server before you install: 5-dimension score breakdown, installs, stars, security level, links

Related MCP server: MCP Servers Search

Scoring model (out of 100, open & reproducible)

Quick start

pip install -r requirements.txt    # installs the MCP SDK (mcp==1.29.1)
python mcp_radar_server.py --stats # print dataset stats
python mcp_radar_server.py         # run as a stdio MCP server

Claude Desktop config (claude_desktop_config.json):

{
  "mcpServers": {
    "mcp-radar": {
      "command": "python",
      "args": ["/absolute/path/mcp_radar_server.py"],
      "env": { "PYTHONUNBUFFERED": "1" }
    }
  }
}

Structure

mcp-radar/
├── mcp_radar_server.py    # MCP server (official MCP SDK / FastMCP)
├── pyproject.toml         # project metadata + pinned mcp==1.29.1
├── requirements.txt       # mcp==1.29.1
├── README.md
├── LAUNCHGUIDE.md         # marketplace onboarding guide
└── data/
    └── mcp_radar_data.csv # 19,947 scored servers

Data refresh

Re-scrape the latest data with mcp_radar_extract.py, then copy the generated CSV into data/.

Roadmap

• Per-server token cost estimation (schema-size proxy)

• Deep rankings for non-Developer-Tools categories (the underserved niches)

• Web leaderboard page (GitHub Pages)

License

MIT