How to build an MCP server people will actually use
· 5 min read
The fastest way to build an MCP server today is to have your AI coding tool do it. Host it at a public URL, sign users in with OAuth if your tools touch their account, and write a clear description for every tool. Then submit it to the ChatGPT and Claude directories.
We'll cover how to build it, the choices popular connectors made, and how to get listed. Not sure you should build one yet? Start with whether your website should be an MCP server.
To see what the popular connectors chose, we read every listing in both stores, compared each store's top 100 with everyone else, and called every server URL:
Build it with your AI coding tool
You don't need to read the whole spec yourself. The MCP project recommends a set of skills that teach your coding agent how to build an MCP server. They ask you a few questions first (what it connects to, who uses it, sign-in, UI), then build to fit. In Claude Code, install them with:
/plugin marketplace add anthropics/claude-plugins-official
/plugin install mcp-server-devUsing Cursor, Codex or another agent? Copy the skill folders into your agent's skills folder. Then run it in your codebase, so the tools match your real data model and API. We share a prompt you can paste in our website post.
If you want to read the code yourself, start with the MCP docs and OpenAI's Apps SDK.
Build a remote MCP server
A local server runs on the user's own computer, like a desktop app. A remote server lives at a URL, like a website. For a store listing, remote is the only real option. Claude's directory no longer accepts local servers on their own, and OpenAI asks for a "stable, public HTTPS endpoint."
So deploying an MCP server is the same as deploying a web service. 21,100 listings (97%) point to a public server URL. Most of the rest are still remote, with a different URL for each customer. Salesforce and Snowflake work this way.
We called all 21,100 URLs. Most asked us to sign in, which is what they should do. But 531 didn't answer, sent back an error, or weren't MCP servers at all. That's about one listing in 40 that a user could install and get nothing back. Check that yours answers before you submit.
MCP server auth: sign in with OAuth
If your tools touch a user's account, use OAuth. It's the "Log in with Google" flow: the user logs in on your site, and the AI app gets a token, never a password. About two in three connectors (65%) use it, a third need no sign-in, and API keys are rare (215 connectors, or 1%).
Several tools do this out of the box, like Stytch Connected Apps, WorkOS AuthKit and Auth0 for MCP, or you can build it yourself. Look for Client ID Metadata Documents, or CIMD, which is the recommended way to build authentication for MCP servers today. Older guides use dynamic client registration (DCR), but a new server should support CIMD.
The top of each store signs in more. All of ChatGPT's top 100 use OAuth, and so do 92 of Claude's:
- ChatGPT: OAuth sign-in
- Claude: OAuth sign-in
- ChatGPT: shows UI
- Claude: shows UI
- ChatGPT: bundles skills
- Claude: bundles skills
We don't think sign-in earns the rank. The leaders are tools people already use every day, like Gmail and Drive, and those need an account (more in what top-ranked connectors have in common). Claude's eight leaders with no sign-in serve public data, like PubMed at #30. If your tools only serve public data, skip sign-in.
How many tools an MCP server needs
A tool is one action the AI can take, like "search messages". The median connector in Claude's top 100 has 26 tools, about double the 12 for everyone else. That gap says more about the apps than the strategy. One Atlassian connector covers both Jira and Confluence, for example.
Many leaders do one thing well:
Indeed ranks #24 in Claude with two tools. So write one tool for each job a user would ask for, and stop there. Every extra tool is one more description the model has to read before it picks.
Spend your own time on those descriptions. They're all the AI reads before it picks a tool, so write them for a model, not a developer.
Annotations, UI and skills
Tool annotations are labels that say if a tool only reads or can destroy something. Both stores now require them, and OpenAI wants all three hints set. 9 in 10 ChatGPT plugins have them. Our MCP security post covers how to label well.
Your tools can answer in plain text, or show a widget in the chat. In Claude, 56% of the top 100 show UI, against 10% of the rest. In ChatGPT the gap is six points, because UI is common across the whole store. Our advice: start with text. Add UI where people need to look before they act, like picking a product or checking a date. (More in our post on UI in chat.)
Skills follow the same pattern as UI, and we explain why in skills over MCP.
Use the current protocol version
The current version is 2026-07-28. Of the 7,534 servers that told us theirs, most are a version behind:
| Protocol version | Servers | Share |
|---|---|---|
| 2026-07-28 (current) | 2,130 | 28% |
| 2025-11-25 | 4,340 | 58% |
| 2025-06-18 | 743 | 10% |
| 2025-03-26 or 2024-11-05 | 321 | 4% |
An older version still works, because clients agree on a version both sides support. But an official SDK keeps you current with no extra work.
Get listed in ChatGPT and Claude
Both stores take submissions in a developer portal, and both review them. For Claude:
- Open claude.ai/directory/manage and choose MCP connector. Any paid plan can do this.
- Paste your HTTPS URL. Your tools sync from the server.
- Give every tool a title and a read-only or destructive hint.
- Pass an automated scan. You're listed as Community by default.
For ChatGPT:
- Verify your identity in your OpenAI organization.
- Upload a plugin ZIP at platform.openai.com/plugins.
- Connect your server and verify its domain.
- Add five positive and three negative test cases, a demo video and a test account without MFA.
- Submit for review, then publish once it's approved.
After you're listed, your rank changes every day. We track ranks for every connector in our directory.
We've built dozens of MCP servers for brands. Reach out if you want help with yours.
Methodology
Data from our daily census of the ChatGPT and Claude stores on November 14, 2026: 21,638 connectors, ranks from each store, and auth, UI and skills from each listing. We called each public server URL to record its status and protocol version. Medians count connectors with at least one tool. More on our method page.
Can you make your website an MCP server?Nov 14, 2026 · 6 min read
ChatGPT plugins are back. Here's what changed since 2023Nov 14, 2026 · 4 min read
Claude connectors: what they are and the ones people useNov 14, 2026 · 4 min read