How to Use an MCP Server with ChatGPT

TL;DR - ChatGPT now supports MCP (Model Context Protocol), letting it connect to external tools the same way Claude Desktop does. - An MCP server exposes tools to any compatible AI client — CRM records, databases, search, custom APIs — via a single standard protocol. - ChatGPT only supports remote HTTP/SSE MCP servers, not local stdio-based ones. Plan your infrastructure accordingly. - Getting set up takes under 20 minutes with an existing hosted MCP endpoint. - OpenHelm's hosted MCP server at mcp.openhelm.ai is remote by design and works with ChatGPT out of the box.

---

Why ChatGPT Users Care About MCP Now

For the first year after Anthropic published the Model Context Protocol in late 2024, MCP was largely a Claude story. Claude Desktop was the flagship client, the ecosystem of integrations grew around it, and most tutorials assumed you were an Anthropic user.

That shifted in early 2025 when OpenAI added native MCP support to ChatGPT. Overnight, the protocol became genuinely model-agnostic — and suddenly very relevant to teams whose AI stack runs on GPT-4o rather than Claude.

If you've been watching Claude users connect their assistant to CRM systems, internal databases, and custom APIs, you can now do exactly the same with ChatGPT. This guide covers how an MCP server for ChatGPT works, how to configure one, and what to avoid.

---

What Is an MCP Server?

An MCP server is a lightweight process that exposes a set of tools to an AI model. In this context, "tools" means functions the model can call during a task: query a database, look up a CRM record, search the web, send a message, trigger a workflow.

Without MCP, connecting an AI model to external tools means custom function-calling code for every integration — a proprietary format that ties you to one provider's API. MCP standardises that interface. A server built to the specification works with Claude, ChatGPT, Cursor, and any other MCP client without modification.

From the model's perspective it's simple: the server publishes a list of available tools and their parameter schemas. The model reads that list, then calls the relevant tools as it works through a task. The server handles execution and returns results. The model never needs to know the implementation details.

For a full breakdown of the protocol itself, see our guide on what is an MCP server.

---

ChatGPT vs Claude Desktop: the Key Difference

The most important practical difference between using MCP with Claude Desktop versus ChatGPT is transport support.

Claude Desktop handles both local (stdio-based) and remote (HTTP with Server-Sent Events) servers. You can run an MCP server process on your laptop, point Claude Desktop at it over a local socket, and it just works — no infrastructure required.

ChatGPT only supports remote MCP servers. The endpoint must be publicly accessible over HTTPS and must speak HTTP/SSE transport. You cannot point ChatGPT at a process running on your local machine.

If you're building a custom MCP server to use with ChatGPT, it needs to be deployed somewhere accessible — a cloud function, a VPS, or a managed platform. A server that only runs locally will never connect.

---

Step-by-Step: Connecting an MCP Server to ChatGPT

This assumes you have a ChatGPT Plus, Team, or Enterprise account (tool integration is not available on free tier).

Step 1: Choose your MCP server

You have three options.

Pre-built community servers. There are over 2,000 MCP servers listed on mcp.so covering GitHub, Notion, Slack, Postgres, Google Docs, and many others. If a server exists for your tool and provides a hosted remote endpoint, this is the fastest path.

Managed platforms. Services like OpenHelm's MCP server at mcp.openhelm.ai give you a hosted remote endpoint with built-in authentication, tool catalogues, and credential vaulting. You connect your tools once through the dashboard; the server handles discovery and execution. Choose this if you need audit trails, access control, or tools that require secure secrets management.

Build your own. Anthropic's open-source MCP SDK supports Node.js and Python. If you need integrations specific to internal systems, you can build a compliant server and deploy it yourself. Our developer documentation covers the full integration pattern.

Step 2: Get your endpoint URL and authentication details

Every remote MCP server has a base URL and an authentication mechanism. The most common patterns are:

• Bearer token — Authorization: Bearer YOUR_KEY header on every request
• OAuth 2.0 — you authorise once and the server refreshes tokens automatically
• No auth — only appropriate for public read-only servers

Make a note of both before you touch ChatGPT's settings.

Step 3: Add the server in ChatGPT

In ChatGPT, go to Settings → Tools (the exact path varies slightly by client version). Select the option to add a custom tool source or MCP endpoint and enter:

1. The server URL (e.g., https://mcp.openhelm.ai/v1)
2. Your authentication credentials
3. A display name you'll recognise

ChatGPT connects to the server, reads the available tools via the MCP discovery handshake, and shows you the full list. Select which tools you want the model to be able to call.

Step 4: Test it

Start a new conversation and explicitly ask ChatGPT to use a tool: *"Use the [tool name] to look up X."*

A working connection will show ChatGPT calling the tool, displaying the result it received, and continuing with that data in context. If the call fails, check authentication headers first, then verify the server is returning Content-Type: text/event-stream for SSE responses.

---

What Can ChatGPT Do Through MCP?

Once connected, ChatGPT can call any tool your server exposes. In practice, that covers:

• CRM lookups — retrieve account or contact records in the middle of a task
• Live database queries — pull current data rather than relying on the model's training
• Document search — vector search across your internal knowledge base
• Workflow triggers — create tickets, send messages, kick off processes
• Custom calculations — business logic that the model would otherwise hallucinate

One important constraint: MCP tool calls are synchronous from the model's perspective. The model calls a tool, waits for the result, and moves on. Long-running jobs (tasks that take minutes) need to return a job ID immediately, with a separate polling tool for status. For genuinely orchestrated, multi-step work — rather than one-off tool calls — you want an agentic workflow platform wrapping your MCP server rather than relying on ChatGPT's chat loop alone.

---

Common Setup Mistakes

Using a local stdio server. ChatGPT simply cannot reach a process on your laptop. The server must be deployed and accessible over HTTPS.

Returning credentials in tool output. Tool responses are visible to the model and included in conversation context. Never return secrets, tokens, or passwords as tool output. Inject them server-side using a credential vault.

Writing vague tool descriptions. ChatGPT chooses when and how to use a tool based on its description and schema. A description like "does stuff with data" produces unpredictable usage. Specific, action-oriented descriptions produce reliable calls.

Ignoring rate limits. In a longer agentic task, ChatGPT might call a tool dozens of times in one session. Build rate limiting and retry logic into your server before you hit production.

---

Frequently Asked Questions

Does ChatGPT support MCP natively?

Yes. OpenAI added native MCP support to ChatGPT in early 2025. It's available on Plus, Team, and Enterprise plans. ChatGPT only supports remote HTTP/SSE MCP servers — local stdio servers are not supported.

Can the same MCP server work with both ChatGPT and Claude?

Yes. A server using HTTP/SSE transport is model-agnostic by design. The same endpoint can serve Claude Desktop, Cursor, ChatGPT, and other MCP-compatible clients simultaneously without modification.

Is there a free MCP server I can use with ChatGPT?

Many community servers are free and open-source, though you'll typically host them yourself. Managed options like OpenHelm's MCP server offer a hosted endpoint with a free tier, at the cost of some customisation flexibility.

What authentication does ChatGPT support for MCP?

ChatGPT supports Bearer token headers and OAuth 2.0. Bearer token authentication is the most straightforward to implement and works reliably across all MCP clients.

How is MCP different from ChatGPT custom GPT actions?

Custom GPT actions predate MCP and use OpenAPI specs tied to the GPT builder. MCP is a separate model-agnostic standard built for programmatic, agent-style tool use. For new integrations, MCP is the better choice — it works across clients and doesn't lock you into the GPT builder format.

---

The simplest path to an MCP-enabled ChatGPT is a hosted server that already understands your stack. OpenHelm's MCP server connects to your existing tools, vaults credentials server-side, and logs every tool call — whether the caller is Claude, ChatGPT, or Cursor.