Key Takeaways
ChatGPT connects to an eligible remote MCP server through its web Plugins flow: you add the endpoint, review the discovered tools, install the plugin, and select it in a chat.
OpenAI states full custom MCP support for Business and Enterprise/Edu on web, while Pro can connect custom servers with read/fetch permissions. Account and workspace policy still applies.
The server needs a public HTTPS endpoint, or a Secure MCP Tunnel if it is private or runs on your machine.
A connected server exposes only its own advertised tools, not every feature of the platform behind it.
Start with a read-only request you can verify. ChatGPT may ask for confirmation on write actions, but you still review the inputs and result.
To use a ChatGPT MCP server, open Plugins on ChatGPT web, choose Add custom MCP server, enter the server's remote endpoint, and complete authentication. Then review the discovered tools, install the plugin, and select it in a chat. Your account and workspace permissions decide whether the option appears.
The Model Context Protocol (MCP) gives ChatGPT a standard way to call a server's tools, but it doesn't hand ChatGPT every feature of the service behind that server. Only the tools the server advertises are available, and your plan and workspace decide whether you can add one at all.
This guide is for SEO pros and technical operators who already have a server URL, or who need to choose between a built-in app and a custom server. I'll cover the permission check, the connection, a first test, and fixes for common failures.
Interface and availability checked November 2026. Connect only servers you trust and review each tool's permissions before use.
Connect an MCP server to ChatGPT: the short answer
On ChatGPT web, the order is: check your access, open Plugins, choose Add custom MCP server, enter the remote endpoint, authenticate, review the discovered tools, then install the plugin and test it in a new chat, as OpenAI's current setup guidance describes. Plan 10 to 20 minutes for a server that's already running. Checking access first matters most, because your plan and role decide which options you see.
Check what you need before connecting
MCP is an open standard for connecting AI applications to external systems. A server exposes tools (actions it can perform) and may also expose resources (data) and prompts. You don't need to code to add a ready remote endpoint; building a server is a separate developer task.
One distinction saves time. A built-in ChatGPT app is a ready-made connection. A custom MCP server is your own endpoint. An MCP app is a server that adds an optional interactive interface. Older tutorials mention Developer Mode and connectors; the current OpenAI flow starts at Plugins.
Have these ready:
ChatGPT web access with a plan and workspace permission that allow custom servers
The full remote endpoint URL, or a Secure MCP Tunnel for a private server
Authentication details, such as OAuth sign-in
The tools or data the server exposes, and one task to test
Add and test the server in four steps
For a remote server that's already running, with credentials and permissions ready, I'd budget 10 to 20 minutes. Building the server, hosting it, and troubleshooting OAuth add time that this estimate leaves out.
1. Confirm access and choose the server
Start with permissions, not the server. OpenAI states full custom MCP support for Business and Enterprise/Edu on ChatGPT web, and Pro accounts can connect custom servers with read/fetch permissions. In a Business workspace, only admins and owners create and publish plugins. In Enterprise/Edu, admins grant permission through RBAC. Founders: that usually means you, or a teammate you've given the admin role.
Then pick a trusted server whose tools match your task. For SEO work, Keytomic's public MCP page lists 59 tools behind one remote endpoint, with account sign-in and no API keys. Those are Keytomic's published claims, so compare them with the tool list ChatGPT shows you. Tip: choose the task before the server.
2. Prepare the endpoint and sign-in
Get the full endpoint URL. ChatGPT connects to remote servers, and OpenAI's developer guide says the server must be reachable through a public HTTPS endpoint or Secure MCP Tunnel. A public endpoint uses streamable HTTP, typically at /mcp, so paste the full path, not just the domain.
Confirm whether sign-in uses OAuth or another method, and keep those details handy. A server on your laptop or private network can't be reached directly; the Secure MCP Tunnel connects it without exposing it to the public internet. Installing a server is separate from adding its endpoint. Warning: don't open a development server to the internet just to test it.
3. Create the custom server connection
Open Plugins in ChatGPT on the web, select the plus button, then choose Add custom MCP server. Enter a name and an optional description. For a public server, provide the HTTPS endpoint; for a private one, choose Tunnel.
Set the authentication method and read the risk warning. Check I understand and want to continue, then select Create as a plugin. Account and workspace controls can change which choices appear, so a missing button usually points to permissions, not a mistake on your side. Tip: name the connection for the job, such as 'SEO data (read-only)', so teammates know what it does.
4. Review tools and run a harmless first request
After creation, ChatGPT shows the tools it discovered. Read each tool name, description, input, and permission, and check they match what the provider promised. Then install the plugin, start a new conversation, type @, and select it.
Ask a narrow read-only question you can check against the source system, such as a count or a short list. Don't start with publishing, deleting, or editing. OpenAI's developer guide recommends MCP Inspector for direct tool tests with representative inputs, including missing identifiers and edge cases. Tip: write down the expected answer before you ask.
Understand what ChatGPT can do with connected tools
In MCP terms, ChatGPT is the client and your plugin points it at a server. Messages follow MCP's JSON-RPC 2.0 base protocol, and the server decides which tools, resources, and prompts exist. ChatGPT sees only what the server advertises.
Based on your request and each tool's name, description, and input schema, ChatGPT picks a tool and fills in the inputs. It can pick wrong, so check the call. A vendor's full product isn't exposed either, only the tools it publishes. A server can run with tools alone; an MCP app adds an interface.
I separate three kinds of work:
Retrieve: read data, such as query rows or an audit list.
Prepare: draft a brief or a proposed change while nothing changes in the source system.
Act: write or modify data. ChatGPT may ask for confirmation, and you review the exact inputs.
Apply MCP to practical SEO and marketing tasks
These examples apply only when your server exposes the matching tool. MCP doesn't connect ChatGPT to Google Search Console, WordPress, Shopify, or Webflow by itself. For a deeper walkthrough, see our Google Search Console MCP workflow, and for what to hand off after approval, these SEO automation workflow examples.
Search Console data: ask for query rows for a named property, date range, and filters, then request the supporting rows before any recommendation.
Technical review: ask for the audit issues with affected URLs, then spot-check two of them in the source tool.
Content briefs: ask for a draft brief built from returned keyword data; nothing is saved until you approve it.
CMS checks: read the current page first, ask for a proposed change as a draft, and leave publishing to a separate, reviewed call.
My validation checklist: property, date range, filters, returned evidence, then approval.
Limit permissions and review security before acting
Treat every connection as a permission boundary. MCP's security best practices recommend minimizing authorization scopes and avoiding broad permissions. OpenAI warns that untrusted servers can introduce prompt injection and other risks, and that ChatGPT may ask for confirmation on write actions. I wouldn't treat that prompt as your review step.
Use this checklist:
Start with read-only tools; add write tools only after the reads check out.
Grant the narrowest scopes and credentials the task needs.
Keep secrets and API keys out of prompts.
Remember the server receives what ChatGPT sends it, so skip sensitive data.
Review exact write inputs and results, and keep a record of what you approved.
Fix common ChatGPT and MCP connection problems
Most failures trace to permissions, the endpoint, sign-in, or the tool list. Match your symptom below and run the one check beside it.
Symptom | Likely cause | Check |
|---|---|---|
No Add custom MCP server option | Plan, role, or workspace policy | Ask an admin to confirm your permission |
Connection fails | Wrong path or unreachable endpoint or tunnel | Verify the full URL, including /mcp where used, and tunnel status |
Sign-in fails or loops | Failed OAuth | Retry authentication and review requested scopes |
Tool list empty or incomplete | Server not advertising tools, or tools changed | Inspect advertised tools in MCP Inspector |
Schema or transport errors | Invalid tool schema or wrong transport | Test edge-case inputs; confirm streamable HTTP |
Timeouts | Slow server or network restriction | Run the same input in MCP Inspector |
One cause is easy to miss: OpenAI says approved workspace tool snapshots don't update automatically. After a server change, an administrator may need to review and refresh the tool list.
How Keytomic fits into connected SEO workflows
Keytomic publishes an SEO-focused MCP server. Its public MCP page lists 59 tools behind one remote endpoint, with account sign-in and no API keys, and describes uses such as Search Console data, site audits, keyword research, and content workflows. That's the difference from a general-purpose server: the tools are scoped to SEO tasks, so the list you review in step 4 reads like an SEO toolkit.
The MCP tool list defines what ChatGPT can do, not the wider platform's feature list. Keytomic's platform covers research, topical planning, content creation, technical audits, CMS publishing, and visibility monitoring. For that bigger picture, see its SEO workflow for marketing teams.
MCP Servers with ChatGPT: Frequently Asked Questions
Does ChatGPT support custom MCP servers?
Yes, on ChatGPT web. Full custom MCP support is stated for Business and Enterprise/Edu; Pro can connect custom servers with read/fetch permissions. Account and workspace policies still apply.
Can ChatGPT connect to a server running on my computer?
Not directly. ChatGPT connects to remote servers, so a local or private server needs OpenAI's Secure MCP Tunnel instead of a public endpoint.
Do I need to code to connect an MCP server?
No, not to add an already-running remote server. You need the endpoint and sign-in details; building or deploying a server is separate developer work.
What is the difference between a ChatGPT app and an MCP server?
A server exposes tools. An MCP app can add an interactive interface, and a built-in app is a separate, ready-made connection rather than a custom server.
Can ChatGPT make changes through MCP?
Yes, if the server exposes write tools and your permissions allow it. ChatGPT may ask for confirmation, but you should review the exact inputs and result.
Why did a tool disappear or stop working after an update?
Check the connection, the tool schema, and whether your workspace's approved tool list needs refreshing. An admin may have to review and refresh it after server changes.
Start with one tool you can verify
My recommendation is to begin with a trusted, read-only tool and one task whose answer you can check. Expand to writes only after you've reviewed the inputs, the access, and the results. The useful outcome is a verified tool call inside an explicit permission boundary, not a promise that everything is automated.
Here's the decision path:
If a built-in app already meets the task, use it.
If the capability is missing and your account and workspace permit it, add a custom MCP server.
If the server is private or local, connect it through Secure MCP Tunnel.
If SEO is your use case, Keytomic's MCP server is one place to start, and its SEO automation for SEO teams page shows where it fits. Pick one read-only task this week, run it end to end, and let the result decide what you connect next.
Resources
Internal resources
External resources
About the author
Salam Qadir, Product & Growth Lead
Salam Qadir is a product owner and growth strategist with 7+ years of experience working across SaaS, product marketing, SEO, and AI search. He focuses on shaping product positioning, developing go-to-market strategies, and building sustainable growth systems for technology products and teams.
His work sits at the intersection of product, marketing, and search, helping companies turn technology into clear market propositions and measurable growth opportunities.
Follow him for practical strategies, insights, and lessons on product, growth, and AI SEO.
