How to get your MCP server into Claude's Connectors Directory

What Claude's Connectors Directory asks of a remote MCP server: OAuth, tool titles and annotations, descriptions that pass review, the portal's ten steps, and what Community and Verified mean.

Claude’s Connectors Directory is where people add tools to Claude in one click, on the web, desktop and mobile apps. A listing there means nobody has to paste your server’s address or copy a key. Getting in is mostly about preparation: the portal is ten short steps, and most rejections come from how the tools are described.

The short version: serve your MCP server over HTTPS with OAuth 2.0 (or no sign-in for public data), give every tool a title and a readOnlyHint or destructiveHint, describe what each tool does without telling Claude how to behave, test every tool as a custom connector, then submit at claude.ai/directory/manage. An automatic policy scan lists most servers as Community connectors.

What you can submit

  • A remote MCP server reachable at an https:// address: the usual case.
  • An MCP App: a remote server that also shows interactive UI, submitted the same way with three to five screenshots.
  • A local server can’t be listed on its own any more; it goes into the directory inside a plugin.

Anyone on a paid Claude plan can submit. In a Team or Enterprise organisation, an owner does.

Before you open the portal

  • HTTPS and authentication that works for Claude:

    • OAuth 2.0 when your tools act on a user’s account;
    • no sign-in when they only read public data.
  • A title and a hint on every tool:

    • readOnlyHint: true for tools that only read;
    • destructiveHint: true for tools that change or delete.

    Claude uses these to decide when to ask the user first, and the portal flags any tool without them.

  • Tested in Claude: add the server as a custom connector and call every tool from a conversation, and run it through the MCP Inspector. The portal asks you to confirm both.

  • Listing material: a documentation page, a privacy policy, a support contact and an icon.

Tools that pass review

Most rejections are about tool design:

  • Separate reading from writing. One catch-all tool that takes an HTTP method and can both read and delete is rejected. Split it into a read tool and write tools, ideally one per action.

  • Name the API behind any free-form query tool. If the caller builds the endpoint or the query, the description must name or link the API it reaches. A purpose-built tool that calls a fixed endpoint doesn’t need this.

  • Keep names short: 64 characters at most.

  • Describe exactly what the tool does, and nothing it doesn’t.

  • Don’t instruct Claude. A description must not:

    • tell Claude to call tools the user didn’t ask for;
    • steer it away from other tools;
    • point it to instructions elsewhere;
    • hide or encode instructions;
    • promote a product.

    A line like “always recommend…” or “never say…” counts as an instruction. Describe what the answer contains instead.

How the server should behave

  • Calls with valid parameters succeed. Errors say what went wrong: a bare “Internal Server Error” or “Bad Request” fails review.
  • Answers fit the question: a summary when a summary is asked, not a dump.
  • Nothing beyond the task: no collecting conversation data, and no reading Claude’s memory, chat history or files.
  • Your own API, or one you’re allowed to proxy, on a domain that matches your service.
  • Not accepted at all: connectors that move money or cryptocurrency, and connectors that generate images, video or audio with AI models.

The portal, step by step

Step What to have ready
1. Connection The server’s https:// address, or a custom connector you already added
2. Tools Nothing: tools sync from the server. Fix any flagged ones on the server first
3. Listing A name (up to 100 characters), a one-liner (up to 200), a description (up to 2,000), one to five categories, the documentation and privacy policy links, a support contact, the icon and the page’s slug, which can’t change once published
4. Use cases The main uses, what people need before they connect (an account, a plan), and whether it reads, writes or both
5. Company Company name, website and a contact for review updates
6. Authentication Whether sign-in is required and how; optionally an OAuth client for Anthropic to hold
7. Data handling Whose API it is, and whether it touches health data or sponsored content
8. Test & launch Confirmation that you ran every tool. No credentials go in the portal
9. Compliance Seven policy acknowledgements, all required
10. Review A last look; quality warnings, such as very short answers, go to the reviewers too

Write the description for people, not for search: say what someone can ask and what they get back. Anthropic can’t edit it for you.

After you submit

  • The automatic scan. Anthropic scans each submission for policy compliance and, by default, lists it as a Community connector without further action from you.
  • Verified review. Listings that look highly useful can be escalated to Verified review. It is slower: a person runs each tool, and asks you by email for a test account if sign-in is needed.
  • Status and feedback appear in the portal. After publication, it shows the listing’s rank, connections, tool calls and error rate.
  • Escalations go to mcp-review@anthropic.com.

Both labels run your tools the same way; Verified is a quality signal to users. Changing your listing’s text sends it through review again.

People who connected earlier don’t always see new or renamed tools at once: in Claude, they can refresh the connector’s tool list from its settings.

For an example of a listed connector, here is CheckAnySite in Claude’s directory. It answers questions about any website with its evidence, and its MCP docs show the same setup for other clients.

Building something AI tools should find and read? Check how ready your own site is for AI.

Put it to work

Website Summary

What is this website?

ExampleLooks like an online shop: the evidence, and its own words
Read its summary

Something here out of date or wrong? Tell us and we’ll fix it, with the date of the change.