Quick answer
MCP (Model Context Protocol) is an open standard for connecting AI assistants to outside apps and data. An MCP server is the small program that exposes one of those apps, such as a calendar or a file store, to the assistant. It is plumbing, not a product you buy.
Key takeaways
- MCP is an open standard; an MCP server is the program on the app's side that lets an assistant read data or take actions there.
- You will usually see it labeled a connector or an integration inside an assistant. You do not need to install a server yourself.
- An assistant connected to an app can act in it, so grant the narrowest access that does the job and keep confirmation prompts on.
- If you only need app A to trigger something in app B on a schedule, a no-code automation platform is the simpler tool.
What is an MCP server?
MCP (Model Context Protocol) is an open standard for connecting AI assistants to outside apps and data, and an MCP server is the small program that exposes one of those apps to the assistant. It is plumbing, not a product you buy. The protocol's documentation describes MCP as "an open-source standard for connecting AI applications to external systems," and compares it to a USB-C port: one standard plug instead of a custom cable for each app. Anthropic's announcement, dated Nov 25, 2024, says "MCP was created at Anthropic by David Soria Parra and Justin Spahr-Summers."
Both big assistants document it, each with its own caveat (pages read October 7, 2026):
- ChatGPT. OpenAI says custom MCP servers "enable you to connect your ChatGPT workspace to external applications." Its limit, on the same page: "custom MCP servers are not developed or verified by OpenAI."
- Claude. Anthropic says people who use Claude "see it as a connector." Its limit: a custom connector is added by entering its URL "with no review by Anthropic."
What does it look like in a real week?
Say you own a small shop and want your AI assistant to draft customer replies using your calendar. In the assistant's settings you find a connector for the calendar and switch it on, choosing read-only access first. The assistant can now see when you are free, so it drafts a reply offering two open slots. When a draft involves sending the message or booking the slot, it stops and asks you first; the specification recommends that a human can deny tool actions, though products vary. After a few weeks you widen the access so it can add events, still with prompts on, and when you stop using it you revoke the connector. This is an illustration, not a description of any named product.
How do the pieces fit together?
Three parts are involved: the host (the AI app you talk to), a client inside it that keeps a connection to one server, and the server on the other side. The architecture page defines the server as "a program that provides context to MCP clients." A server can offer three kinds of things:
- Tools: actions the assistant can run, such as creating an event or searching records.
- Resources: data it can read, such as file contents or database records.
- Prompts: reusable templates that help structure a request.
The architecture page says "MCP server" refers to the program that serves context data, "regardless of where it runs." A local server runs on your own computer. A remote server is run on the app maker's platform, and you meet that kind inside an assistant: choose the connector, sign in, approve what it may touch.
Do I need to know this, or can I ignore it?
For a typical owner the answer is no: you do not need to install or run an MCP server. You will meet MCP as a connector or integration switch. The practical question is whether you want your assistant to reach this app, and with what permissions. The documentation says end users gain assistants that "can access user data and take actions on the user's behalf when necessary." Reading your calendar is low risk. An assistant that can also send email or delete records is a different decision.
You can ignore MCP entirely when:
- your assistant already does what you need with text you paste in;
- the job is fixed and repeats the same way, such as "when a form arrives, add a row and send a confirmation" (a scheduled automation fits better; see no-code automation platforms for small business);
- you cannot narrow what the assistant is allowed to do in the app.
What is the difference between an API and an MCP server?
An API is an app's own doorway for other software; an MCP server presents an app's features in a shared format so any compatible assistant can use them. Apps usually design their APIs separately. MCP gives the assistant one consistent way to ask what an app can do and then call it. The specification says servers expose tools that "can be invoked by language models." A connector is only as capable as what its builder chose to expose.
The table is our own simplified explanation, not a statement from the specification.
| Direct API integration | MCP connector | No-code automation | |
|---|---|---|---|
| Typical use | A developer's software calls the app in a fixed way | An assistant calls tools during a conversation (other software can call MCP servers too) | A visual flow runs on a trigger you set |
| Setup you do | Hire or build | Usually: switch on a connector, approve access | Build or copy a flow |
| Main limit or risk | Needs a developer to build and maintain | Broad access if permissions are wide, plus prompt injection | You build and maintain each flow |
What should I check before connecting an assistant to an app?
Check what the connector can do, give it the narrowest permission that works, and keep the confirmation prompts on. The specification says "there SHOULD always be a human in the loop with the ability to deny tool invocations." The word "should" makes it a recommendation to assistant builders, not a guarantee in every product.
- Name the job in one sentence, such as "read my calendar and draft replies." If you cannot, do not connect yet.
- Pick read-only first if the connector offers it.
- Read the permission screen. The MCP security guidance recommends a "progressive, least-privilege scope model."
- Keep confirmations on for anything that sends, edits, pays or deletes.
- Prefer connectors from the app's own maker where one exists. This is our own advice, not a quotation. For context, the specification says clients should treat tool annotations as untrusted unless the server is trusted.
- Review and revoke in the connected-apps list now and then.
The risk behind this list is prompt injection: OWASP says "A Prompt Injection Vulnerability occurs when user prompts alter the LLM's behavior or output in unintended ways," so text in an email or web page can try to steer an assistant that has access. MCP's security guidance tells clients that offer one-click setup of local servers to warn users that MCP servers run with the same privileges as the client, so install a local server only from a developer you trust.
Will this change, and what does it cost?
The standard is still moving, so check dates; the protocol itself is open-source and not something you pay for. The documentation pages are versioned 2026-07-28 (read October 7, 2026), and one page marks sampling, a feature that servers can request from the client, as deprecated as of that version. Costs come from the assistant plan and the app you connect; see each vendor's own pricing page. For what assistants that act on your behalf cost, see AI agents for small businesses. For a simpler start with no MCP, see the one-weekend automation stack.
How was this page made?
Definitions come from the MCP documentation and specification (version 2026-07-28) and Anthropic's announcement; statements about ChatGPT and Claude come from OpenAI's and Anthropic's own documentation; security statements from the specification, its security guidance and OWASP. This page was drafted with Claude, an AI model made by Anthropic, and checked against each vendor's published material on October 7, 2026. Anthropic created MCP, so the origin statement is attributed to its own announcement. Confirm connector details in each vendor's current help pages.
Next step: open the AI assistant you already use, find its connectors or integrations list, and read what the first one you care about is allowed to do before you switch it on.
FAQ
Do ChatGPT and Claude use MCP?
Yes, both document support. The MCP documentation names Claude and ChatGPT as AI applications that can connect through MCP. OpenAI's developer documentation, in a section on risks, says custom MCP servers "enable you to connect your ChatGPT workspace to external applications." Its caveat on the same page: "custom MCP servers are not developed or verified by OpenAI, and are third-party services that are subject to their own terms and conditions." Anthropic's documentation says "An MCP server gives Claude access to your product or data, and people who use Claude see it as a connector." Its caveat: a custom connector is one that a user or an organization Owner adds by entering its URL, "with no review by Anthropic." (Pages read October 7, 2026.)
Is MCP free to use?
The protocol's own documentation calls MCP "an open-source standard for connecting AI applications to external systems." That describes the standard, not any product built on it. What an assistant plan or a connected app costs is set by the vendor, so check its pricing page.
Is my data sent to the app’s maker?
It depends on where the server runs, and the documentation does not promise an answer for every product. The MCP architecture page says "MCP server" means the program that serves context data, "regardless of where it runs", and gives two examples: a local server on your own computer and a remote server run on the app maker's platform. Check the app's privacy page and the permission screen for what the connector can read.
Can I run my own MCP server?
Yes, the protocol is open and anyone can build one, but that is developer work and outside this page. If you want a connection nobody has built yet, a freelancer or a no-code automation platform may be the simpler route.
How do I know if an app I use offers a connector?
Look in two places: the app's own integrations or help page, and the connector list inside your assistant. If neither shows one, a no-code automation is the usual alternative.