Screen an MCP server before you wire it into your agent

Before you paste an MCP install snippet into Claude Desktop or Cursor, you usually have one question: does this tool’s description look honest, or does it hide instructions you would not want an agent to follow?

An advisory screen reads the tool description (and related fields) for integrity issues — hidden directives, contradiction, overclaim. On mcpindex the public screen emits REVIEW or UNVERIFIED at v1. ALLOW and DENY exist in the verdict contract but are not produced by the public semantic screen today; a clearing ALLOW waits on a behavioral conformance milestone.

How to use it in practice: find the server on mcpindex (search or the per-server page), read the trust panel, or paste a description into /screen. Treat UNVERIFIED as fail-closed — not permission to proceed. Treat REVIEW as “a human should look,” not “cleared for autonomous use.”

A screen is not certification, not a badge of safety, and not a substitute for an in-path gate. Screens help at install time. Gates help when the live contract changes after install. Publishers who want a live advisory badge for a README can embed the SVG from their server page; the badge tracks the same advisory record the site shows.

Start here: https://mcpindex.ai/screen and https://mcpindex.ai/search