Get Your MCP Server Into Claude's Connectors Directory

You submit a remote MCP server to Claude's Connectors Directory through Anthropic's developer portal at claude.ai/directory/manage: select Submit new, choose MCP connector, and work through ten steps [1]. Anthropic then scans the submission automatically and, by default, lists it as a Community connector with no further action from you [1].

Everything below is checked against Anthropic's documentation on 20 October 2026 [1][2]. Anthropic moved its directory FAQ from the help center into the Claude developer docs [3], so a guide that links to the old help article is the older one.

TL;DR:

The server must be remote and reachable over https:// [1]. A local server cannot be submitted as a connector on its own; it goes inside a plugin [1].
Every tool needs a title and a readOnlyHint or destructiveHint [1]. The portal flags tools missing them when it syncs your tool list.
Authenticated services use OAuth 2.0 [1]. Public data can use no authentication at all.
Have a test account ready. Reviewers need credentials for a fully populated account [2].
Any paid Claude plan can submit [1].

Before you open the portal

Anthropic's own checklist [1], in the order it matters:

A public HTTPS URL. The first portal step asks you to paste it, or to pick a custom connector you already added to Claude.
Authentication that Claude's client can complete. OAuth 2.0 if your tools act on a user's account, none for public data [1]. No separate approval or Verified status is needed for auth to work [1].
Tool annotations on every tool: a title, plus readOnlyHint: true for read-only tools and destructiveHint: true for tools that modify or delete data [2]. These decide permissions inside Claude: read-only tools can run without asking each time, destructive tools always prompt [2].
You have run every tool yourself, through the MCP Inspector or as a custom connector in Claude. The Test & launch step asks you to confirm it [1].
Listing materials: a documentation URL, a privacy policy URL, a support contact and an icon [1].
A reviewer test account with real-looking data in it [2].

What each portal step asks for

Step	What you provide [1]
1. Connection	The server's `https://` URL, or a custom connector you already added. Choose multiple URLs or a URL pattern if each user connects to their own address
2. Tools	Nothing to type: tools, prompts and resources sync from the server, grouped by read-only or write. Fix flagged tools on the server first
3. Listing	Name (100 characters), one-liner (200), description (2,000), one to five categories, docs URL, privacy URL, support contact, icon, and a URL slug that is permanent once published
4. Use cases	The main use cases, what a user needs before connecting, and whether it reads, writes or both
5. Company	Company name and website, and a contact for review updates
6. Authentication	OAuth with Dynamic Client Registration, Client ID Metadata Documents, or Anthropic-held client credentials; a custom connection; or none
7. Data handling	Whose API it is (yours, proxied with permission, or a third party's), and whether it handles health data or sponsored content
8. Test & launch	Step-by-step reviewer access, with every link and credential
9. Compliance	Seven policy acknowledgments, all required
10. Review and submit	A final check. Quality warnings, such as very short answers, go to the review team with the submission

Progress saves in your browser as you go [1]. The description you write in step 3 is yours alone: Anthropic cannot edit it [1].

Which sign-in to pick in step 6

Dynamic Client Registration is what most guarded MCP servers already offer. In our registry census we read the authorization server metadata of 169 servers that asked for credentials: 162 offered Dynamic Client Registration and 33 accepted Client ID Metadata Documents [4]. If your server is among the few that offer neither, the remaining option is client credentials that Anthropic holds for you.

What gets a connector sent back

Anthropic's review criteria list the common rejection reasons [2]:

One tool that both reads and writes. A catch-all api_request tool with a method parameter is rejected. Split reads from writes, and ideally writes into create, update and delete [2].
Missing annotations, or tool names over 64 characters [2].
Descriptions that steer Claude instead of describing the tool: telling it to call other software, to fetch instructions from elsewhere, or to promote a product [2].
Tools that fail with valid input, or that return a bare "Internal Server Error" or "Bad Request" with no detail [2].
Calling an API that is not yours and that you do not proxy with permission. The server's domain should match your service [2].
Unsupported use cases: moving money or crypto, and generating images, video or audio with AI models. Diagrams, charts and UI mockups are allowed [2].

A free-form query tool also needs to name the API it calls, or link to its docs, in its description [2].

After you submit

The automatic scan lists a passing submission as a Community connector [1]. Anthropic may later escalate a connector it judges highly useful to Verified review, in which reviewers run each tool; that happens automatically and needs nothing from you [2]. Status and reviewer feedback appear in the same portal, and Anthropic's submission page gives the review team's email address for escalations [1].

Check the public half first

The directory's first step connects to your URL from Anthropic's side, not from your laptop. Our MCP connection test makes that same kind of outside call and shows what comes back: a full handshake with the tool list, a 401 from a server guarding itself, or the exact reason it failed. To wire the server into Claude yourself before submitting, see add a remote MCP server to Claude Desktop or Claude Code.

A directory listing puts your server in front of people browsing Claude's directory. A launch week puts it in front of people watching what shipped this week. Submit it to MCP Hunter: it is free, a person reviews every submission, and when we can reach your server the listing shows what it returned, with the date we asked.