MCP tool description checker
This MCP tool description checker reads your server's tool list and checks each name, description and annotation against the rules Anthropic and OpenAI publish for connectors and plugins.
This tool is in development.
The checker tests these rules for each tool.
| Check | Rule | Source |
|---|---|---|
| Tool name is 64 characters or fewer | Hard rule (Claude) | Anthropic review criteria |
| Name pairs a domain and an action, with a shared prefix across related tools | Best practice | OpenAI metadata guide; Anthropic "Writing tools for agents" |
| Description says what the tool does and when to call it ("Use this when…") | Hard rule and best practice | OpenAI submission guidelines; Anthropic review criteria |
| Description does not tell the model how to behave or promote products | Hard rule (Claude) | Anthropic review criteria |
title is set; readOnlyHint or destructiveHint is set (OpenAI also asks for openWorldHint) | Hard rule | Both |
| Read and write tools are split; no catch-all tool | Hard rule (Claude) | Anthropic review criteria |
Parameters in inputSchema have clear names, descriptions, examples and enums | Best practice | OpenAI metadata guide; Anthropic |
Rules checked 3 October 2026.
Every MCP tool needs a clear name, description and annotations.
Name
Use a namespace prefix and an action, like
crm_contacts_search. Keep it under 64 characters.Description
Start with "Use this when…". Say what the tool returns and when not to call it.
Parameters and annotations
Write
user_id, notuser, and give each parameter a description, an example and an enum where values are fixed. Settitle,readOnlyHintanddestructiveHinton every tool, andopenWorldHintfor ChatGPT.
Frequently asked questions.
Why does Claude reject my connector?
Common reasons are tool names over 64 characters, missing title or hints, catch-all tools, and descriptions that tell Claude how to behave. A missing privacy policy also means rejection.
Do tool descriptions affect which server the model picks?
Yes. Anthropic and OpenAI both say the model reads your tool names and descriptions to decide when to call a tool. Find your server in the MCP servers registry.
Is my JSON stored?
No. The checker runs in your browser.
Does this check server.json?
No. To check your listing and server.json fields, use the AI app listing grader. Fix your full listing with app listing optimization.