Guides · Checked November 15, 2026
How to get your MCP server into ChatGPT, Claude and Cursor
Publishing to the official MCP Registry takes one command, and nobody reviews it. The directories inside the AI apps work differently. Each one has a review, some check who you are, and most keep you waiting for weeks with little news. These guides cover what each app asks for, why servers get rejected, how long people actually wait, and where to get help.
At a glance
| App | Who can submit | What you submit | Typical wait |
|---|---|---|---|
| ChatGPT | Anyone with a verified identity at OpenAI | A plugin ZIP that points at your remote server | About 2 weeks, sometimes months |
| Claude | Anyone on a paid Claude plan | A remote server URL, or a plugin repo on GitHub | Hours to days since the new portal |
| Cursor | Mostly companies, open source only | A plugin in a public Git repo | 1 to 2 weeks per Cursor staff, often a month or more |
| Perplexity | Companies, by proposal | A proposal form | No timeline and no guarantee |
| GitHub Copilot and VS Code | Servers already in the official MCP Registry | An onboarding request | Weeks, or no answer |
| Google Gemini | Developers Google lets in, in phases | A remote server, in Connected Console | Not published |
| Microsoft 365 Copilot | Verified Microsoft partners who own the service | A certification package | Limited capacity during the preview |
| Meta Muse | Businesses | A remote server, on the Muse Connector Platform | Not published |
None of these apps commit to a review time. The waits above are what developers report on forums and on reviewtimes.fyi, which collects them.
What every reviewer checks
The rules differ per app, but the same problems get servers rejected everywhere. Fix these before you submit anywhere:
- A public, stable endpoint. Streamable HTTP over HTTPS, on a domain you own. No tunnels, no localhost, and no redirect to another host. Check that your firewall or bot protection lets reviewers through: Anthropic, for example, connects from
160.79.104.0/21. - A test account a stranger can use. Email and password, already filled with sample data. No two-factor codes, SMS, magic links or sign-up step. "Can't connect" is the first of the common rejections OpenAI lists.
- Annotations on every tool. Set
title,readOnlyHintanddestructiveHint(andopenWorldHintfor ChatGPT) explicitly, and make them match what the tool really does. Sending a message or queuing a job counts as a write. Deleting or overwriting is destructive, even when it can be undone. - Separate read and write tools. A catch-all tool that takes an HTTP method or runs arbitrary operations gets rejected by both OpenAI and Anthropic.
- Descriptions that describe instead of steer. Write them like API reference docs. No "ALWAYS call this first", no instructions about other tools, no marketing copy and no prices.
- OAuth done by the book. For ChatGPT and Claude: PKCE with S256, protected-resource metadata, and dynamic client registration or a Client ID Metadata Document. Without a token, answer
401(not403) with aWWW-Authenticateheader. Microsoft is the exception, and supports neither dynamic registration nor PKCE. - A privacy policy that matches your responses. List every field your tools return and make sure the policy covers it. Strip internal IDs, trace IDs, timestamps and debug fields.
- Clear errors and small responses. Specific error messages instead of "Internal Server Error", and responses well under the limits (Claude cuts results off at about 150,000 characters).
- Test prompts that work on their own. Reviewers run your example prompts word for word, on a clean account, on web and mobile. A prompt that only works after another one fails.
- Your own service. ChatGPT, Claude and Microsoft all reject connectors to an API you don't own or have permission to proxy.
Check your server before you submit
- MCP Inspector, the official tester. Both OpenAI and Anthropic tell you to call every tool with it first:
npx @modelcontextprotocol/inspector@latest. - MCPJam Inspector, open source. Emulates ChatGPT and Claude, runs ChatGPT-style test cases, and grades a server against Claude's rules with
mcpjam readiness check claude. - Alpic Beacon renders your app in the real ChatGPT and Claude and gives a readiness verdict for each. Needs an Alpic account, and only works on servers without sign-in.
- mcpready, free:
npx mcpready preflight <url> --target claude(or--target openai). - Dawloom's preflight script, free, comes with a 30-point ChatGPT checklist quoted from OpenAI's docs.
- McpStamp Check and Transica's directory check, free web checks that score a server per app.
- Suncly: a person tests your server the way a reviewer would, for $100 to $200, with a report in one or two days.
Companies that submit for you
Plenty of agencies build MCP apps. Only a few prepare the submission, send it and see it through review, and fewer show results. Prices are from their own sites.
| Company | What they do | Price | Proof |
|---|---|---|---|
| Alpic Studio | Builds the app, prepares the submission packages and works directly with the AI labs | Not public | Case studies: Facile.it, AirHelp and Snipp, live in ChatGPT and Claude |
| Transica | Builds the app from your API, records the demo, fills in the ChatGPT and Claude forms and supports you through review | From $1,500, plus hosting from $19 a month | One ChatGPT approval, October 2026 |
| Invokary | Submits from your accounts to Claude, ChatGPT, Cursor and the MCP Registry, with two rejection rounds included | Audit $750, launch from $9,500 | Only its own server so far |
| McpStamp | Fixes what blocks review, writes the listing and submits to Claude, ChatGPT, Gemini and the MCP Registry | $4,900, audit $490 | No case studies |
| Modern Mustard Seed | Files the Claude and ChatGPT submissions | Quoted after a call | Its own connectors in Claude's directory |
Two platforms write the paperwork but leave the submitting to you: Alpic's Submission Helper drafts the ChatGPT and Claude forms from your live server ($50 a month), and Manufact's submission pack generates listing copy, test cases, a demo video and the ChatGPT ZIP ($5 each).
Is it worth it?
A listing is mostly about being found, not about access: in almost every one of these apps, users can already add any remote server by URL. And being listed doesn't mean being shown. OpenAI puts only plugins it picks on the directory's main pages, and everyone else is found by exact name or direct link (OpenAI). Claude ranks its directory by usage, and 47 of the 50 most popular connectors in August 2026 came from Anthropic's partners (Nicolas Sitter's census).
The directories are filling up fast. ChatGPT's went from about 2,300 plugins in August 2026 (Node8) to more than 13,000 in November (OpenApps), and Claude's lists close to 10,000 connectors, most of them Community listings (mcptop). So get listed, but plan to bring your own users too: a listing makes you installable in one click, but it rarely sends traffic on its own.
The guides
- ChatGPT: plugin ZIP, identity check, 5 positive and 3 negative test cases, and review that can take months.
- Claude: one portal for connectors and plugins, Community listings within days, and Verified review by invitation.
- Cursor: a hand-picked marketplace for open-source plugins, and the self-serve cursor.directory.
- Perplexity: a proposal form for companies, and custom connectors in the meantime.
- More AI apps: GitHub Copilot and VS Code, Google Gemini, Microsoft 365 Copilot, Meta Muse, Grok and others.
- Connectors vs plugins: what connector, plugin, app and extension mean in each app, and which one to submit.