自作MCPサーバー その10 #AI - Qiita
@ohisama@github

Self-made MCP Server Part 10

  • Self-made MCP
  • AI
  • ClaudeDesktop
  • ChatGPT
0
Last updated at 2026-11-12Posted at 2026-11-12

Overview

I got curious about MCP servers, so I looked into it.
I asked the AI.
What is a custom connector?

What is a Custom Connector

"How to develop your own custom connector (remote MCP server)" is the explanation.
The essence of a custom connector is an "MCP server" hosted online on the internet. Claude accesses the specified URL (https://...) via HTTP communication and calls the functions of the program you created.

2. Specific Creation Method

The flow of creating a custom connector and connecting it to Claude is as follows:

  1. Writing the MCP server program
    Using the official SDK for Python or TypeScript (Node.js), define the functions (tools) you want the AI to execute.
  2. Hosting it on the internet
    Prepare an HTTPS URL so that it can be accessed from Claude's server.
  3. Registering the URL in Claude's interface
    Enter the issued URL from "Add Custom Connector" to connect.

2. Specific Creation Method

Step 1: Creating the Server with the Official SDK

Using the official MCP SDK published by Anthropic (Python or TypeScript), you can easily create it with just a few dozen lines of code.
• Specifying the communication method (transport)
If you're running it only on your PC, use "stdio". However, if you're using it as a custom connector (remote MCP), you need to start the server in the "Streamable HTTP (or SSE)" format that allows communication via the web.
• Defining the tools
Create functions such as "get ○○" or "calculate △△", and add explanatory text (hints for when the AI should use the function).

Step 2: Hosting it on the Internet via HTTPS

Place the created program on the cloud and prepare a URL accessible from the outside.
• Easy hosting services (hosting platforms)
Using PaaS services such as Cloudflare Workers, Vercel, Render, Railway, and Fly.io, along with AWS Lambda, you can easily generate an HTTPS URL just by connecting it to GitHub.
• Testing methods during local development
Even before deploying to the cloud, you can assign a temporary HTTPS URL to your server running on a Windows PC using tools like ngrok or Cloudflare Tunnel, allowing you to test the connection immediately from Claude.
• Implementation of authentication (OAuth, etc.)
For personal use, you can implement simple authentication, but for others to use, you'll need to incorporate OAuth 2.0 (such as Google Login) (Koi Launch also uses this mechanism).

Step 3: Register the MCP server as a custom connector with Claude

Once the URL is ready, you can link it on Claude's side interface.

  1. From the bottom icons (or settings) in Claude's left menu, open Customize → Connectors.

    • Click Add → Add custom connector.
  2. Enter a name (optional) and the MCP server URL (https://...) prepared in Step 2 and save it.

  3. When you turn on the connector via the "+" button in the chat screen, you will be able to use the custom functionality you created.

  4. Official resources helpful for development

When actually writing code, the following official documentation and tools are most helpful.
• Model Context Protocol official documentation (modelcontextprotocol.io)
Tutorials for building servers using Python (mcp / FastMCP) and TypeScript (@modelcontextprotocol/sdk) are available.
• MCP Inspector (official debugging tool)
This is an official tool that allows you to test whether the tool is properly recognized and whether errors occur on your browser before connecting it to Claude.
• Official Cloudflare Workers remote MCP template
There is an official template that allows you to deploy an OAuth-enabled remote MCP server in a few minutes, and it is often used by individual developers.

Yes, you can also connect and use your own MCP server (custom connector) in ChatGPT.

Since MCP is a common specification, you can register and reuse the same remote MCP server URL (https://...) you created for Claude on the ChatGPT side as well.

Connection steps for ChatGPT (developer mode)

When connecting a self-created MCP server to ChatGPT, enable "developer mode (Developer mode)" in the settings and register it.

  1. Open the settings screen
    From the profile icon at the bottom left of the ChatGPT screen, open "Settings" and select "Apps & Connectors" (or "Connectors" depending on the environment).
  2. Turn on Developer Mode
    Open "Advanced settings" at the bottom of the screen and turn on the switch for "Developer mode".
  3. Register your custom MCP server
    When you return to the previous screen, the "Create" button will be displayed, so click it and enter the following information.
    • Name: Any name (example: My Custom MCP)
    • MCP Server URL: The URL of the published server (https://...)
    • Authentication Method: Choose "No authentication" or "OAuth" etc.
  4. Call in chat
    Open a new chat window, and select the registered connector from the "+" button (or @ mention) in the input field, allowing you to use your custom functionality.

Points to note when using Claude and ChatGPT together

• One server is enough
If you have one remote MCP server (Streamable HTTP compatible) created for Claude's custom connector, you can use it with both Claude and ChatGPT by registering the same URL for both.
• ChatGPT-specific UI extension (Apps SDK)
If you just need to exchange data or run tools, it's fine to use the standard MCP. However, if you want to display custom operation buttons or card-type UI within the ChatGPT interface, you can also incorporate the specifications of OpenAI's "Apps SDK".
• Note on usage plans
You may need a paid plan (Plus / Pro / Team / Enterprise) to use Developer Mode (adding custom MCP) in ChatGPT.

There are many well-known custom connectors (remote MCP servers) that add advanced features such as search, development, automation, and design to AI.

1. Connectors that enhance AI capabilities and information collection

• DeepWiki (DeepWiki)
A remote MCP that allows AI to deeply understand open-source code on GitHub and provide explanations and answers. It is widely used by engineers when researching new libraries and tools.
• Exa Search / Firecrawl / Apify
Connectors that allow AI to directly perform highly accurate information extraction or website-wide scraping (data collection) instead of regular web searches.

• Context7
A popular server that allows AI to refer to the "latest official documentation" of various programming libraries and frameworks in real time. This helps prevent AI from writing code based on outdated versions.

2. Business automation and no-code integration type

• Zapier MCP (Zapier)
A tool that converts the automation tool "Zapier", which can connect to thousands of business apps, into an MCP. By connecting just one custom connector, you can delegate various actions such as LINE notifications, spreadsheet updates, and Trello operations to AI.

• Dify / Make(Self-hosted / Custom Workflow)
It's common to issue the complex business flow you created with no-code AI development tools like "Dify" or automation tools like "Make" as a remote MCP URL, and then call it from Claude or ChatGPT.

3. Development, Cloud, and Database-related

• Supabase / Neon
An official remote MCP that allows you to create and search for database tables, and update data using natural language in chat. It works well with app development and is sometimes used together with app generation tools like Koi Launch.
• Cloudflare / Vercel / Netlify
A popular connector for developers that allows direct commands from chat for deploying (publishing) websites and apps, changing server settings, and checking logs.
• Sentry(Sentry)
It's often used to have AI automatically read error logs from web services or apps, identify the cause, and suggest fixes.

4. Design, Payment, and Other Unique Connectors

• Figma / Canva
You can have AI convert design data structure into code or instruct it to generate or edit images and presentation materials.
• Stripe / PayPal
An official remote MCP that allows you to operate payment data and check sales status, and create billing links directly from chat.

Directory site to find famous custom connectors

You can search for remote MCP servers that can be used by simply copying the URL, categorized by genre, on the following specialized directory sites.
• Remote MCP Directory(remote-mcp.com)
A listing site specialized in "remote type (custom connector)" MCP servers that can be used by simply entering a URL.
• MCP.so / Awesome Remote MCP Servers(GitHub)
A representative directory that collects MCP servers published by developers around the world, and features a large number of remote-compatible URLs.

※Note that ultra-popular tools such as Gmail, Google Drive, Notion, and Slack are already standardly included in the "official directory" of Claude and ChatGPT, so you can connect them with one click from the screen without having to manually enter URLs.

End.

Go to the list of users who liked