How to Use MCP with ChatGPT (Complete 2026 Guide)

ChatGPT doesn't know anything about you when you open a new chat. Every time, you re-paste the same context: who you are, what you're working on, the decision you made last week.

MCP fixes that. You connect ChatGPT to a server that already knows. ChatGPT searches it, reads it, writes back to it. Your context compounds instead of resetting.

This is the practical guide. What MCP is, the two ways to connect ChatGPT to one, the real examples that show why it's worth doing, and the troubleshooting steps you'll actually need.

What MCP is, in one paragraph

MCP (Model Context Protocol) is an open standard introduced by Anthropic in November 2024 and adopted by OpenAI in 2025. It lets an LLM call external tools through a uniform protocol. Search this, read that, create the other thing. Instead of every app shipping a custom plugin or GPT, any app that speaks MCP can be wired into any AI client that speaks MCP. ChatGPT, Claude, Cursor, and a growing list of others all read from the same servers.

If you want the longer version, read What is MCP?. For why it's better than custom GPTs, see MCP vs Custom GPTs.

The two ways to connect ChatGPT to an MCP server

There's the easy way and the developer way.

Easy way: Apps Directory. ChatGPT publishes a directory of MCP-backed apps. One click connects them. No URLs, no developer mode, no settings spelunking. If the server you want is in there, use this path.

Developer way: custom MCP servers. Any remote (HTTPS-reachable) MCP server can be added to ChatGPT through developer mode. This is how you connect anything that isn't (yet) in the Apps Directory, including servers you build yourself.

Both end up at the same place: ChatGPT can call the server's tools mid-conversation. The difference is how much you click to get there.

Path A: Connect via the ChatGPT Apps Directory

If your server is published in the directory, this is the whole setup:

1. Open the app's page in ChatGPT (for Hjarni, that's the Hjarni app in ChatGPT).
2. Click Connect.
3. Sign in to the underlying service if it requires OAuth.

That's it. Open a new conversation and ChatGPT can already call the app's tools. The Apps Directory route works on web; mobile support varies by app and rollout.

When this works, it really is one click. The trade-off is that the directory is curated. Not every MCP server is in there. Hjarni went live in the directory in April 2026; the launch post shows what the connection flow looks like end to end.

Path B: Connect a custom MCP server with developer mode

For anything not in the directory, like your own server, an early-stage tool, or an internal company server, use developer mode.

Step 1: Enable developer mode

1. Open ChatGPT and go to Settings → Apps.
2. Scroll to Advanced settings.
3. Toggle Developer mode on.

Developer mode is currently in beta and available on Plus, Pro, Business, Enterprise, and Education plans, on the web. It supports remote MCP tools, including write and modify actions, subject to ChatGPT's confirmation flow. If you don't see the toggle, your plan, workspace settings, or platform is the reason.

Step 2: Create an app from the MCP server

1. In developer mode settings, click Create app.
2. Paste the MCP server URL. The server must be a remote (HTTPS-reachable) MCP server. ChatGPT does not support local stdio servers (the kind that start with command: node or command: npx). For Hjarni that's https://hjarni.com/mcp. For your own server, whatever you've deployed.
3. If the server uses OAuth, complete the auth flow when prompted.
4. Review the tools the server exposes. Toggle on the ones you want ChatGPT to be able to call.

Tip: leave write tools off until you've tested read tools. It's the cheap way to feel confident before letting an LLM modify anything. Only connect MCP servers you trust. Untrusted servers can introduce prompt injection or data exposure risks, especially with write tools enabled.

Step 3: Use the app in a conversation

1. Start a new conversation.
2. Open the + (plus) menu and choose Developer mode.
3. Select the apps you want active for this conversation.

The "select per conversation" step is the one most people miss. The app is enabled globally in settings, but you still pick it from the + menu before each chat that needs it. Until you do, ChatGPT acts like the server doesn't exist.

For the docs version of this walkthrough, see How to connect ChatGPT to an MCP server.

Verifying it works

Start a new conversation and ask ChatGPT to do something only the server could do.

What notes do I have in Hjarni about my Sweden trip?

Read my "About me" note and use it as context for this conversation.

Search my knowledge base for anything I've written about MCP.

If the connection is live, you'll see ChatGPT call a tool (it's visible in the response) and answer with content from the server. If it answers without a tool call, the app isn't enabled for this conversation. Go back to the + menu.

Real examples: what MCP with ChatGPT actually unlocks

The setup is the easy part. The reason to do it is the workflow that opens up. A few that I run every week using Hjarni's MCP server:

Stop re-explaining the project. I have a "Hjarni context" note that summarizes the product, the stack, the current priorities. ChatGPT reads it before writing copy, before suggesting features, before answering anything project-specific. I haven't pasted product context into a chat in months.

Capture without breaking flow. Mid-conversation I'll say "save the part where you compared MCP to plugins as a note in Research." ChatGPT writes the note, tags it, drops it in the right folder. Next week, when I ask a related question in a new chat, it pulls that note back. The synthesis tax gets paid once.

Cross-LLM context. Both Claude and ChatGPT read from the same Hjarni knowledge base. I plan in Claude Code, refine in ChatGPT, query from my phone. Same notes everywhere. No syncing. This is the LLM wiki pattern without the local-only friction.

Marketing and SEO loops. I keep my copywriting guide, my SEO priority list, and my zero-click pages list in Hjarni. ChatGPT reads the guide, rewrites titles, drops the rewrites back as a note. I review and ship. The whole marketing loop lives inside ChatGPT and one knowledge base.

The point isn't any single example. It's that ChatGPT stops being a forgetful assistant and starts being one that remembers.

Troubleshooting

The four problems that account for almost every "it doesn't work" report:

Developer mode toggle is missing. You're on a plan or platform that doesn't have it yet. Developer mode is currently web-only and on paid plans. The Apps Directory works more widely, with mobile support varying by app.

App is "connected" but ChatGPT ignores it. You enabled it globally in settings but didn't pick it from the + menu for the current conversation. Pick it. New conversations don't inherit per-conversation tool selections.

OAuth never completed. The first connection to a server like Hjarni opens an auth flow. If you closed the tab, the connection is half-finished. Remove the app and add it again. That re-triggers the flow cleanly.

MCP server URL has a typo or trailing slash. ChatGPT doesn't always give a useful error. Double-check the URL exactly matches what the server documents. For Hjarni it's https://hjarni.com/mcp. No trailing slash, no /api.

If none of those, check that the specific tools you want are toggled on inside the app's tool list. ChatGPT can only call what you've allowed.

Which MCP servers should you start with

There's a long and growing list. The honest short list for ChatGPT users:

• Hjarni: knowledge base and notes, in the Apps Directory. The case I know best, because I built it.
• Notion, Linear, Sentry, Stripe, GitHub: all have official MCP servers and most are in the directory.
• Anything you build yourself: MCP is open. If your team has a tool with a JSON API, wrapping it as an MCP server is a weekend project.

The rule of thumb: connect the server that holds the context you keep re-pasting. That's the one with the highest payoff. For most people, that's notes. For some, it's a CRM, a ticketing system, or a code search tool.

Where to go next

You've got ChatGPT calling tools. That's the floor, not the ceiling. The next move is making the server worth calling.

• If you specifically want Hjarni with ChatGPT: the Hjarni for ChatGPT guide walks through what each tool does and example prompts.
• If you're using Hjarni: the Knowledge Wiki template sets up a structure that ChatGPT can navigate well: sources, topics, open questions, AI instructions per folder.
• If you also use Claude: connect Claude through MCP and both clients read the same notes.
• If you're new to the pattern: Give Claude Long-Term Memory and How to Build an LLM Wiki with Claude and MCP cover the same ideas from the Claude side. The setup is mirror-image; the workflow is the same.

---

ChatGPT with MCP is the version of ChatGPT that knows your project. The setup is ten minutes. The payoff is every conversation after.

Connect Hjarni to ChatGPT. Free to start, no credit card.