Before you publish an MCP server, run 10 checks. Most servers fail at least three — and the failures are invisible until an agent picks the wrong tool, hallucinates an argument, or silently drops your server on connect. This is the checklist we built mcp-probe to enforce, distilled to what actually breaks in the wild.
TL;DR — A publishable MCP server connects cleanly, names tools unambiguously, describes every argument, validates inputs, and ships install metadata. The single most common failure is thin tool descriptions: even the five official Anthropic reference servers cap at 60/100 on description quality.
Why "it works in Inspector" isn't enough
MCP Inspector answers "does my server connect and list tools?" That's necessary, not sufficient. The agent doesn't experience your server the way you do in a UI — it experiences your tool descriptions and schemas as text in a context window, and it picks tools by reading them. A server can pass Inspector and still be functionally unpublishable because the model can't tell your tools apart.
So the pre-publish question isn't "does it run?" It's "is it publishable?" — will a real agent, with no docs and no human in the loop, use it correctly?
