Why connector submissions get rejected on Claude, ChatGPT and Cursor

Each directory publishes the rules its reviewers check, and many of them are mechanical: a missing tool title or hint, a name over 64 characters, a description that tells the model what to do, one tool that both reads and writes, test credentials that don't work, or, on ChatGPT, the wrong number of test cases. You can check every one of these before you submit. Below are the rules we track for Claude, ChatGPT and Cursor, each linked to the requirement on our platform pages, where its source and the date we last checked it are.

We don't publish rejection rates: no directory shares them, and we won't guess. What follows is what the platforms say they check.

Claude

Connector and plugin submissions go through automated checks and a safety scan, and the developer portal shows the review status and the changes Anthropic asks for. The connector rules:

• Tool metadata. Every tool has a title and the applicable hint: readOnlyHint for read-only tools, destructiveHint for tools that modify or delete data.
• Names. Tool names are 64 characters or fewer.
• Descriptions. They say what the tool does and when to use it, and must not tell Claude how to behave. "Use this tool to look up an invoice by number" passes; "Always call this tool first" does not.
• Mixed tools. A single tool that mixes safe and unsafe HTTP methods, such as a catch-all api_request tool, is rejected. Read and write actions go in separate tools.
• Test credentials. A connector submission includes test credentials for a fully populated account. An empty account gives reviewers nothing to check.
• Plugins. For a plugin bundle, the first organization to submit a repository folder holds that listing, so submit your own before anyone else does.

ChatGPT

OpenAI's plugin submission has more paperwork, and some of it is checked by count:

• Hints with reasons. Every MCP tool sets readOnlyHint, openWorldHint and destructiveHint to match what it does, with a justification for each.
• Test cases. The submission includes exactly five positive and three negative test cases. Not four, not six.
• Reviewer access. Reviewer test credentials work without MFA, SMS or email confirmation, or private-network access.
• Domain token. A domain verification token is served at /.well-known/openai-apps-challenge on the MCP server's host or a parent host.
• Eligibility. Projects with EU data residency cannot currently submit MCP plugins.

After publication, OpenAI periodically fetches your tools: a changed tool goes live only once it passes automated checks, so the same rules apply to every later change.

Cursor

Every plugin is reviewed by hand before it is listed, from a public Git repository:

• Manifest. Cursor accepts a root plugin.json or a .cursor-plugin/plugin.json, and plugin names are lowercase kebab-case.
• Licenses. Open-source components must use permissive licenses, such as MIT, BSD or Apache 2.0. GPL, AGPL and LGPL are not allowed, so check your dependencies too.

Check these before you submit

1. List your server's tools with the MCP Inspector and read each one as a reviewer would: title, hints, name length and description wording.
2. Look for any tool that takes an HTTP method or an arbitrary endpoint as an argument, and split it into separate read and write tools.
3. Sign in with the test credentials from a fresh browser profile and confirm the account has real data and no second factor.
4. For ChatGPT, count the test cases, open the domain token URL, and write a justification for each tool's hints.
5. For Cursor, check the manifest's name and every license in the repository.

If you were already rejected

Read the reason the platform gives against the rule lists above, fix the cause rather than the symptom, and resubmit. If the fix means redesigning tools or moving the server to the current MCP spec, that is the Connector Upgrade. Every requirement for each directory, with its source and check date, is on our Claude, ChatGPT and Cursor pages.