All posts

Publishing an MCP server to the ChatGPT and Claude directories

· Dev

We published the Memocore MCP server (shared memory for AI agents) with 29 tools in the official ChatGPT and Claude directories. ChatGPT took 28 days and one rejection. Claude took 37 days with no feedback at all, and it may have been stuck for no good reason — more on that below. This post covers what the directories require, how review works, where we tripped up, and what happens when you change tools after publishing.


Why get into a directory at all. Without it, the user connects the server manually: copies the address, enters the connector name. From the directory it's one button and a screen where they allow access. The directory also brings a little traffic, especially in ChatGPT: the assistant itself recommends your service to users who are looking for a solution to their task. Over the last month, 27 people signed up from ChatGPT, 7 over the last week.


What the server needs before submitting

The requirements of the two directories are almost the same: ChatGPT, Claude.


  • A remote MCP server over HTTPS. A local server can't be submitted to the Claude directory on its own, only inside a plugin.
  • Sign-in via OAuth 2.0: the client has to find on its own where to authorize (addresses in /.well-known) and register on its own (Dynamic Client Registration); the code exchange uses PKCE.
  • Annotations on every tool: title (specifically inside annotations, not at the top level of the tool description), readOnlyHint, destructiveHint, idempotentHint, openWorldHint. By the MCP spec they're optional, but it's better to set all five explicitly. We got our first rejection in ChatGPT exactly because of one optional annotation — idempotentHint.
  • Privacy policy, documentation and a support address on the site.
  • A test account with data that the reviewer will log into on their own. If sign-in is by a code from an email, like ours, the reviewer needs a fixed code: they don't have access to your email.
  • For the listing: icon, short description, categories and example prompts. Claude asks for at least three working examples. OpenAI currently asks for five prompts where the tools should trigger, three where they shouldn't, and a demo video.
  • Only for ChatGPT: domain verification. At /.well-known/openai-apps-challenge the server has to return a token from the OpenAI dashboard.

ChatGPT

Memocore page in the ChatGPT directory


We submitted under the old moderation scheme: the whole submission was checked by a person. Before submitting, we set readOnlyHint, destructiveHint and openWorldHint on the tools, set up domain verification and a demo account with a fixed sign-in code.


We submitted on August 6. On August 21, 15 days later, we got a rejection. There was one remark: none of the tools had idempotentHint. We added it to all tools. true — where a repeated call changes nothing: reading, deleting, access settings. false — where every call leaves a trace: creating records and editing memos, because every edit adds a new version to the change history.


We sent the fixed version the same day. On September 3, Memocore appeared in the directory: 28 days from the first submission to publication. Our logs show the reviewer did log into the test account.


Claude

Memocore page in the Claude connectors directory


Before, Claude accepted directory submissions only from organizations on the Team and Enterprise plans. On September 25, Anthropic opened submissions to all paid plans through the portal claude.ai/directory/manage. We submitted on October 7.


In Claude, the tool name in the listing is taken from annotations.title, not from the title field at the top level of the tool description. If only the second one is filled, the portal marks the tools as unnamed.


The form in the portal has ten steps: server address, listing, use cases, company details, authorization, data handling, test account and a list of requirements you need to confirm. In the notes field, we explained in advance everything they could pick on. For example, our tool descriptions tell the agent to follow a project's instruction — formatting rules written by the project's owner. From the outside that can look like third-party instructions to the model, so we explained where they come from. We created a separate test account for the Claude reviewer.


The submission sat in In review status for 37 days, and no emails came during that time. On November 13, I was about to write to support, but before that I pressed Sync tools in the portal: since submitting, we had added 3 more tools. A couple of seconds later, an email came saying the connector was approved.


Whether it's a coincidence or the submission was stuck because the tool list no longer matched what was submitted, I don't know. If you changed tools after submitting, press Sync tools right away.


Judging by the data, no one ever logged into the test account. According to the docs, Anthropic checks submissions automatically and by default publishes them as Community, and a person only looks at some of the submissions. Looks like ours went through without a manual review.


Updates after publishing

ChatGPT

Recently the moderation changed: changes to existing tools are now checked by automation first. If it finds nothing, the change is available to users right away. If it finds something, the change waits until you send it to manual review with the Appeal button. New tools went to manual review every time for us. Manual review usually takes a few days, and all that time automatic updates are on pause.


In our experience, the check most often flags openWorldHint. Even if the tool doesn't go anywhere outside your service, the check may decide it should be true. If false is correct, you need to explain why in the appeal. For us, the check flagged update-project-memo, and here it was right: editing a memo with a public link changes a public page, so we set it to true.


Claude

We haven't rolled out new tools since publishing yet. Small changes to existing tools were applied automatically, without review.