CustomGPT.ai Blog

How to Connect an AI Tool to a CustomGPT.ai Hosted MCP Server

June 25, 2026

16 min read

Current image: any llm aware tool + cusomgpt hosted mcp server

Introduction

To connect an AI tool to a CustomGPT.ai hosted MCP server, open your CustomGPT.ai project, copy your Project ID and MCP token, then add the hosted MCP server URL or JSON configuration to your AI tool’s MCP settings. The URL follows this format: https://mcp.customgpt.ai/projects/<PROJECT_ID>/sse?token=<TOKEN>. After saving and restarting the tool, you can ask questions that are answered from your project’s knowledge base.

A hosted MCP server is a remote server that CustomGPT.ai runs for you, so MCP-aware AI tools can reach your approved project content over a standard protocol. Connecting through MCP lets the tool give grounded answers from your private project data instead of relying only on the model’s general knowledge.

This guide is written for developers, AI builders, product teams, customer support teams, and business users testing MCP tools.

Key Takeaways

A CustomGPT.ai hosted MCP server lets MCP-compatible AI tools connect to your project’s knowledge base over a standard protocol.
You need a CustomGPT.ai account, a project with uploaded or connected content, your Project ID, and your MCP token before you start.
MCP details are available in the dashboard under Deploy then MCP Server (Beta).
The direct connection URL format is https://mcp.customgpt.ai/projects/<PROJECT_ID>/sse?token=<TOKEN>.
Some tools connect directly to the remote MCP server, while others use a local helper such as supergateway.
Treat your MCP token like a password. Never commit it to GitHub, logs, screenshots, or shared chats.
Use the hosted MCP server when your tool already supports MCP, and use the API when you are building a custom app or automation.

What Is a Hosted MCP Server?

A hosted MCP server is a remote server, managed by a provider, that exposes context or tools to AI applications through the Model Context Protocol (MCP). MCP is an open standard, introduced by Anthropic, that defines how AI tools request external context and capabilities in a consistent way.

Think of MCP as a universal plug. Instead of building a custom integration for every AI tool, you connect through one standard interface, and any MCP-aware tool can use it. The hosted part means you do not run or maintain the server yourself. The provider hosts it, and your tool connects to it over the internet.

In the CustomGPT.ai case, the hosted MCP server gives MCP-compatible AI tools a standard way to reach the content inside a CustomGPT.ai project.

Summary: A hosted MCP server is a managed remote endpoint that lets MCP-aware AI tools connect to external content through one standard protocol, without you running the server.

Why Connect an AI Tool to CustomGPT.ai with MCP?

Connecting an AI tool to CustomGPT.ai through MCP lets the tool access a CustomGPT.ai project’s approved knowledge base instead of relying only on the model’s general training. The result is grounded answers based on your own content.

There are three main benefits. First, answers can draw on private project content the model was never trained on, such as your documentation, policies, and product data. Second, answers are grounded in retrieved source material, which helps reduce hallucinations. Third, the same project knowledge becomes reusable across any MCP-aware tool you connect, so you set up content once and reach it from multiple tools. This grounding is the same retrieval-augmented generation approach that powers a CustomGPT.ai project.

Summary: MCP connects an AI tool to your CustomGPT.ai project so it can answer from private, approved content rather than general model knowledge alone.

What You Need Before You Start

Before connecting an MCP-aware tool to CustomGPT.ai, gather the following:

A CustomGPT.ai account.
A CustomGPT.ai project with uploaded, crawled, or connected content.
Your Project ID.
Your MCP token.
An AI tool that supports MCP or external MCP servers.
The tool’s MCP settings panel or configuration file.

If your project has no content yet, add documents or connect a data source first. The MCP connection can only answer from content that exists in the project.

Where to Find Your CustomGPT.ai MCP Details

Your MCP connection details live in the CustomGPT.ai dashboard. To find them:

Log in to CustomGPT.ai.
Open your project.
Go to Deploy, then MCP Server (Beta).

From that screen, copy your Project ID, your MCP token, and any JSON configuration shown for helper-based tools. Keep the token private and only paste it into your own AI tool’s settings.

Two Ways to Connect: Direct Connection vs Helper Connection

There are two common ways to connect an MCP-aware tool to the CustomGPT.ai hosted MCP server. A direct connection has the tool reach the remote server itself. A helper connection uses a local helper such as supergateway to forward requests to the remote server.

Connection type	How it works	Best for	What you usually enter	Example tools
Direct connection	The AI tool connects directly to the CustomGPT.ai hosted MCP server over the internet	Tools that support remote or URL-based MCP servers	The hosted MCP server URL with your Project ID and token	Any tool that accepts a remote MCP server URL
Helper connection	The AI tool runs a local helper such as supergateway that forwards requests to the hosted server	Tools that expect a command-based or local MCP server	A JSON configuration that launches the helper with your server URL	Tools that use a command-based MCP config file

Summary: Use a direct connection when your tool accepts a remote MCP URL. Use a helper connection when your tool expects a command-based local server and needs supergateway to bridge to the hosted server.

Direct Connection Setup

Follow these steps to set up a direct connection:

Open your AI tool’s MCP or external server settings.
Choose add server or add MCP server.
Enter the hosted MCP server URL.
Use this URL format: https://mcp.customgpt.ai/projects/<PROJECT_ID>/sse?token=<TOKEN>
Replace <PROJECT_ID> with the Project ID from CustomGPT.ai.
Replace <TOKEN> with the MCP token from CustomGPT.ai.
Save the connection.
Restart or refresh the AI tool if required.
Test with a question that should be answered from your CustomGPT.ai project content.

Helper Connection Setup with Supergateway

Follow these steps to set up a helper connection using supergateway:

Open the MCP Server section in the CustomGPT.ai dashboard.
Copy the JSON configuration shown for helper-based tools.
Open your AI tool’s MCP configuration file or settings.
Paste the JSON configuration.
Save the file.
Restart the AI tool.
Test the connection with a knowledge-base question.

Note: The exact configuration format depends on the AI tool. Follow the tool’s MCP documentation and use the JSON block provided by CustomGPT.ai when available. For example, see the Cursor MCP docs or the Windsurf Cascade MCP docs for tool-specific configuration paths.

Example MCP Configuration

This is a generic example using placeholders only. Replace the placeholders with your own values from the dashboard.

{
  "mcpServers": {
    "customgpt": {
      "command": "npx",
      "args": [
        "supergateway",
        "--sse",
        "https://mcp.customgpt.ai/projects/<PROJECT_ID>/sse?token=<TOKEN>"
      ]
    }
  }
}

Warning: Never publish or commit MCP tokens to GitHub, public docs, logs, screenshots, or shared chat messages. Treat the token as a secret.

How to Test Your MCP Connection

After connecting, test the link with prompts that should pull from your project knowledge base:

“Answer using the CustomGPT.ai project knowledge base.”
“What does our documentation say about [specific topic]?”
“Summarize the policy from the connected knowledge base and cite the relevant source if available.”
“If the answer is not in the connected project, say that you do not know.”

A successful response will reflect content that exists in your CustomGPT.ai project, not generic model knowledge. If the tool can reference your specific documents, policies, or product details, the connection is working. If it answers vaguely or ignores your content, revisit the setup and troubleshooting steps.

Common Troubleshooting Issues

Problem	Likely cause	How to fix it
Tool cannot find the MCP server	URL is missing, malformed, or the tool does not support remote MCP	Recheck the URL format and confirm the tool supports remote or helper-based MCP
Invalid token	Token is wrong, expired, or regenerated	Copy a fresh MCP token from the dashboard and update the connection
Wrong Project ID	The ID does not match the intended project	Copy the correct Project ID from the project’s MCP Server screen
Server URL copied incorrectly	A character was dropped or extra spaces were added	Paste the full URL again and remove any stray characters or line breaks
Helper command not installed	supergateway or its runtime is not available	Install the helper and confirm the command runs in your environment
AI tool needs restart	Settings were not reloaded after saving	Fully restart or refresh the AI tool, then test again
Content not uploaded or indexed	The project has no usable content yet	Add and index documents or connect a data source in CustomGPT.ai
Query is too vague	The prompt does not point at project content	Ask a specific question that targets your uploaded content
Token exposed or compromised	The token was shared or committed somewhere public	Regenerate the token immediately and update every connection that used it

Security Best Practices for MCP Tokens

Your MCP token controls access to your project content, so protect it carefully. For broader platform controls, see CustomGPT.ai security and AI for compliance. Follow these practices:

Treat MCP tokens like passwords.
Use project-specific access so a token only reaches the intended project.
Do not paste tokens in public places such as repositories, forums, or shared screenshots.
Rotate or regenerate tokens if you suspect they were exposed.
Limit who can access MCP setup details inside your team.
Remove old connections when they are no longer needed.

Hosted MCP Server vs API: Which Should You Use?

The hosted MCP server and the API both connect AI tools to your CustomGPT.ai project, but they fit different needs.

Option	Best for	Requires coding	Typical use case
Hosted MCP server	Connecting MCP-aware AI tools quickly	Little to none	Linking a tool that already supports MCP to your project knowledge base
API	Building custom interfaces, apps, and automations	Yes	Programmatic workflows, custom UIs, and integrations you build yourself

Use the hosted MCP server when your AI tool already supports MCP and you want a simpler connection. Use the API when you are building a custom interface, app, automation, or programmatic workflow.

Best Practices for Better Answers Through MCP

The quality of MCP answers depends on the quality of your project content. To get better results:

Upload clean, authoritative content.
Remove outdated or duplicate documents.
Use clear document titles so content is easy to retrieve.
Test with real user questions.
Ask the AI tool to answer from the connected knowledge base.
Review answers for groundedness against your sources.
Keep the CustomGPT.ai project content updated as your information changes.

How CustomGPT.ai Helps

CustomGPT.ai helps teams create AI agents from approved business content and connect that knowledge to MCP-aware tools through a hosted MCP server. This can reduce the need to build and maintain a custom retrieval layer, server infrastructure, and tool-connection workflow from scratch.

In practice, you upload or connect your content to a project, then expose it to compatible AI tools through the hosted MCP server. The same project knowledge becomes reachable from any MCP-aware tool you connect, which keeps your content in one place while making it usable across tools. This pattern fits customer service assistants and enterprise teams that need grounded answers from approved content.

Conclusion

To connect an AI tool to a CustomGPT.ai hosted MCP server, create or open a project, copy the Project ID and MCP token, add the hosted MCP URL or JSON configuration to the AI tool, restart if needed, and test with questions from the project knowledge base. Choose a direct connection when your tool supports remote MCP URLs, or a helper connection with supergateway when your tool expects a command-based server. Keep your token private, keep your content clean and current, and the connection will return grounded answers from your project.

FAQ

What is a hosted MCP server?

A hosted MCP server is a remote server, managed by a provider, that lets MCP-compatible AI tools connect to external content or tools through the Model Context Protocol. The provider runs and maintains the server, so you do not host it yourself. With CustomGPT.ai, the hosted MCP server gives MCP-aware tools a standard way to reach the content inside a CustomGPT.ai project over the internet.

What is Model Context Protocol?

Model Context Protocol, or MCP, is an open standard that defines how AI tools connect to external context, data, and tools in a consistent way. It acts like a universal plug, so an MCP-aware tool can connect to any MCP server without a custom integration. CustomGPT.ai uses MCP to let compatible AI tools access a project’s knowledge base through a hosted MCP server.

How do I connect an AI tool to a CustomGPT.ai hosted MCP server?

To connect an AI tool to a CustomGPT.ai hosted MCP server, open your project, copy your Project ID and MCP token, and add the hosted MCP server URL or JSON configuration to your tool’s MCP settings. Use the URL format https://mcp.customgpt.ai/projects/<PROJECT_ID>/sse?token=<TOKEN>. Save the connection, restart the tool if needed, and test with a question that should be answered from your project content.

What do I need before connecting an MCP-aware tool?

Before connecting an MCP-aware tool, you need a CustomGPT.ai account, a project with uploaded, crawled, or connected content, your Project ID, and your MCP token. You also need an AI tool that supports MCP or external MCP servers, plus access to that tool’s MCP settings panel or configuration file. If your project has no content yet, add documents or connect a data source before testing the connection.

Where do I find my CustomGPT.ai MCP token?

You find your CustomGPT.ai MCP token in the dashboard. Log in, open your project, then go to Deploy and select MCP Server (Beta). That screen shows your Project ID, your MCP token, and any JSON configuration for helper-based tools. Copy the token and keep it private. Only paste it into your own AI tool’s settings, never into public repositories, logs, or shared messages.

What is the CustomGPT.ai hosted MCP server URL format?

The CustomGPT.ai hosted MCP server URL follows this format: https://mcp.customgpt.ai/projects/<PROJECT_ID>/sse?token=<TOKEN>. Replace <PROJECT_ID> with your project’s ID and <TOKEN> with your MCP token, both available in the dashboard under Deploy then MCP Server (Beta). Tools that support remote MCP servers can use this URL directly, while command-based tools may use it inside a helper configuration.

What is the difference between a direct MCP connection and a helper connection?

A direct MCP connection has the AI tool reach the CustomGPT.ai hosted MCP server itself using the server URL. A helper connection uses a local helper such as supergateway to forward requests from the tool to the hosted server. Use a direct connection when your tool accepts a remote MCP URL, and a helper connection when your tool expects a command-based or local MCP server.

What is supergateway in MCP setup?

Supergateway is a local helper that bridges a command-based AI tool to a remote MCP server. Some tools expect to launch an MCP server with a command rather than connect to a URL. In that case, the helper runs locally and forwards requests to the hosted server. You configure it inside the tool’s MCP configuration file, pointing it at the CustomGPT.ai hosted MCP server URL.

Can Claude connect to a CustomGPT.ai hosted MCP server?

Claude is an MCP-aware AI tool, and MCP-compatible tools can connect to a CustomGPT.ai hosted MCP server using either a direct URL connection or a helper connection. The exact steps depend on the specific Claude product and its MCP settings, so follow that tool’s official MCP documentation. Use the CustomGPT.ai server URL or JSON configuration from the dashboard, and confirm support in the tool’s own docs.

Can Cursor or Windsurf connect to a hosted MCP server?

MCP-compatible coding tools can connect to a hosted MCP server using a direct URL or a helper connection, depending on how each tool implements MCP. The exact configuration format varies by tool, so follow the official MCP documentation for the tool you use. Use the CustomGPT.ai server URL or JSON configuration from the dashboard, and verify MCP support in that tool’s own documentation before setup.

Should I use MCP or the CustomGPT.ai API?

Use the hosted MCP server when your AI tool already supports MCP and you want a simpler connection with little or no coding. Use the CustomGPT.ai API when you are building a custom interface, app, automation, or programmatic workflow. MCP is best for connecting existing MCP-aware tools to your project, while the API is best for custom development you control end to end.

How do I test if my MCP connection works?

To test an MCP connection, ask a question that should be answered from your project content, such as “What does our documentation say about [topic]?” A working connection returns answers grounded in your CustomGPT.ai project, not generic model knowledge. You can also prompt the tool to say it does not know when content is missing. If it references your specific documents and policies, the connection is working.

What should I do if my MCP connection fails?

If your MCP connection fails, check the server URL, Project ID, and token for typos or extra characters first. Confirm the token is current and that your project has indexed content. For helper connections, make sure supergateway is installed and the command runs. Restart the AI tool to reload settings, then test again with a specific question. Regenerate the token if you suspect it was exposed.

Is it safe to share an MCP token?

No, you should not share an MCP token. The token grants access to your project content, so treat it like a password. Never commit it to GitHub, paste it into public docs, logs, screenshots, or shared chats. Use project-specific access, limit who can view MCP setup details, and regenerate the token immediately if it is exposed or no longer needed.

How does CustomGPT.ai use MCP?

CustomGPT.ai uses MCP to connect MCP-aware AI tools to a project’s approved knowledge base through a hosted MCP server. You upload or connect content to a project, then expose it to compatible tools using the server URL or a helper configuration. This lets tools return grounded answers from your content and reduces the need to build a custom retrieval layer and server infrastructure yourself.

Related Resource:

CustomGPT.ai hosted MCP for Trae See how Trae can use CustomGPT.ai as a private knowledge source through Model Context Protocol.
Connect WhatsApp to a chatbot See how WhatsApp Business Platform works with a CustomGPT.ai chatbot for support, lead capture, and appointment workflows.
Knowledge base chatbot setup guide See the steps for turning approved business content into a chatbot that gives grounded, useful answers.
Create a chatbot free Walk through the steps to choose a platform, design conversation flows, train your bot, and deploy it on your website.

Priyansh Khodiyar

Priyansh is a Developer Relations Advocate at CustomGPT.ai who writes deeply researched technical content on RAG APIs, AI agent development, and cloud-native tools.