What Is MCP? The Model Context Protocol in 9 Plain Answers
What is MCP? The Model Context Protocol explained in plain English: how it links AI to your apps, who supports it, the security risks and whether you need it.

MCP, the Model Context Protocol, is a shared standard that lets an AI assistant connect to other software, such as your calendar, accounts package or online shop, and do things there. Anthropic released it in November 2024, and OpenAI, Google, Microsoft and Amazon now support it. The official documentation calls it "a USB-C port for AI applications". You rarely see the letters MCP: you see a button called connectors, apps or plugins.
You ask your AI assistant what last month's Stripe refunds came to, and it says it cannot see your Stripe account. A colleague asks the same question and gets a number, because someone clicked "connect" on a list of apps. Behind that button sits MCP. Until it existed, every AI product needed its own custom link to every tool, and most links never got built. Now a company such as Notion or Canva builds one connection and it works in Claude, ChatGPT and others. By December 2025 Anthropic counted more than 10,000 active public MCP servers.
Below: how we checked this guide, MCP in one table, 9 plain answers, the connections a small business might use, the security record and a short checklist before you connect anything.
Prefer not to wire tools together yourself? The Jobbit agent builds and hosts the tool or automation you describe. Free plan, 100 credits a day, no card.
How we checked this guide
- The official specification and documentation, read on 18 November 2026 at modelcontextprotocol.io, including its security guidance.
- Announcements from the organisations involved: Anthropic, the Linux Foundation, Microsoft, Google and Amazon, plus TechCrunch's reporting of OpenAI's statements.
- Each connector checked on the vendor's own developer page: Stripe, Notion, Slack, HubSpot, Canva, Zapier, GitHub, Cloudflare, Google Drive and Shopify.
- Security incidents from the researchers and outlets that published them. Several are from security vendors, which have an interest in the problem looking large.
- Left out: which ChatGPT plans allow custom connections. OpenAI's help page would not load and third-party guides disagree.
MCP at a glance
| Question | Answer | Source |
|---|---|---|
| What it stands for | Model Context Protocol | Official documentation |
| Who created it | Anthropic, announced 25 November 2024 | Anthropic |
| Who runs it now | The Agentic AI Foundation, part of the Linux Foundation, since 9 December 2025 | Linux Foundation |
| Foundation's top-tier members | AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, OpenAI | Linux Foundation |
| Current version | 2026-07-28 | Official specification |
| Release rhythm | Every six months from March 2027 | MCP blog, 11 November 2026 |
| What you see in Claude | Connectors | Claude help centre |
| What you see in ChatGPT | Apps and plugins | OpenAI, TechCrunch |
| Public servers | More than 10,000 by December 2025 | Anthropic |
| Cost of the standard | Free and open source | Official documentation |
The 9 plain answers
1. What does MCP actually do?
It gives AI applications one standard way to reach outside systems. The official introduction says: "Just as USB-C provides a standardized way to connect electronic devices, MCP provides a standardized way to connect AI applications to external systems." A printer maker does not build a different plug for every laptop. A software company no longer builds a different AI link for every assistant.
2. How does it work, without the jargon?
Three parts. The host is the AI app you use, such as Claude or ChatGPT. The server is a small program run by the tool you want to reach, such as Notion. The client is the piece inside the host that holds the connection. A server can offer three things: tools, which the documentation defines as "executable functions that AI applications can invoke to perform actions", resources, which are data to read, and prompts, which are reusable templates.
3. What is the difference between local and remote?
A local server runs on your own computer and usually serves one app, for example to read files in a folder. A remote server runs on the provider's systems and serves many customers over the internet. For a non-technical owner, remote is the one that matters: you click connect, sign in to the other service in your browser and approve access, the same way you would let an app use your Google account.
4. Who supports it?
Almost everyone. OpenAI's chief executive wrote in March 2025: "People love MCP and we are excited to add support across our products." OpenAI's apps inside ChatGPT, launched in October 2025, are built on it. Microsoft documents MCP in Copilot Studio and Windows. Google Cloud and Amazon Web Services publish their own MCP tooling. In December 2025 Anthropic donated the protocol to a new foundation under the Linux Foundation so that no single company owns it. By October 2026 that foundation had 305 member organisations.
5. Where do I see it as an ordinary user?
- Claude: under Connectors. Custom connectors are available on every plan, and the help centre says free users are limited to one.
- ChatGPT: as apps, with launch partners including Booking.com, Canva, Spotify and Figma.
- Gemini: as connected apps. Adding your own is limited to adults in the United States on personal accounts, in English, according to Google's help page.
- Microsoft 365 Copilot: as Copilot connectors, some of which fetch live data through MCP.
6. Is MCP an app I install or pay for?
No. It is a standard, like Bluetooth. You pay for the AI assistant and for the service you connect, and the connection itself is usually free. Zapier counts MCP use against your plan's task allowance, which is the kind of detail to check.
7. Is it the same as an API, or an automation tool?
An API is a service's own way of letting programs talk to it, different for every service. MCP is a common wrapper around those, written for AI. Automation tools such as Zapier run fixed recipes: when this happens, do that. With MCP the AI decides which tool to use for the request in front of it. The two overlap, and Zapier offers its 9,000 apps through MCP. Our comparison of Zapier, Make, n8n and AI agents covers when each fits.
8. What about A2A and the other names?
A2A, the Agent2Agent protocol, lets one AI agent talk to another. Its own site puts the difference in a line: "MCP connects an agent to tools, APIs, and data. A2A connects it to other agents. Use both." AGENTS.md is a plain instructions file for coding agents. All three now sit under the same foundation.
9. Does a small business need to care?
You need to know two things. First, when you choose software, a working connector to your AI assistant is now a feature worth asking about. Second, every connection is a door, and the security section below explains how to keep it narrow. You do not need to read the specification.
Connections a small business might actually use
Each is confirmed on the vendor's own developer page.
- Stripe. An official server. Actions such as refunds are marked for human confirmation, and Stripe warns users to take care when combining it with other servers.
- Notion. Search your workspace, create and edit pages and manage databases.
- Slack. Workspace administrators approve which AI apps may connect.
- HubSpot. Its remote server became generally available to all accounts on 13 April 2026.
- Canva. Open to any assistant since October 2026. Each person signs in with their own Canva account.
- Zapier. A bridge to more than 9,000 apps, on every plan including free.
- Google Drive. Available in a developer preview, with Google's own warning that the server "can read, modify, and delete data".
- GitHub and Cloudflare. Official servers for people who manage code or websites.
- Shopify. Shop connections exist but changed in 2026, so follow Shopify's current documentation.
The security record
MCP gives an AI real access to real accounts. The specification itself admits that it "cannot enforce these security principles at the protocol level". These incidents are documented.
- Hidden instructions in tool descriptions. In April 2025 Invariant Labs showed that a malicious server can hide instructions in the description of a tool, which the AI reads and the user usually never sees. They called it tool poisoning.
- A poisoned issue on GitHub. In May 2025 the same team showed an attacker could hijack a user's agent through a malicious issue on a public project. They called it an architectural problem, not a bug in GitHub's code.
- A fake package that copied emails. In September 2025 The Register reported a malicious copy of an email provider's MCP server that quietly added a hidden recipient to outgoing mail. The real provider had nothing to do with it.
- A critical flaw in a popular helper. In July 2025 JFrog disclosed CVE-2025-6514 in a widely used connection tool, rated 9.6 out of 10, later fixed.
- A cross-customer leak. Asana took its MCP server offline for nearly two weeks in June 2025 after a bug that could have exposed data between customers, The Register reported.
The researcher Simon Willison describes the dangerous mix as a "lethal trifecta": access to private data, exposure to untrusted content and the ability to send information out. An AI holding all three can be tricked into leaking. The UK's National Cyber Security Centre advised in August 2026 that agents should hold "only the permissions it needs". Our guide to prompt injection explains the underlying attack.
There is also a cost. Anthropic's own engineers wrote in November 2025 that tool descriptions take up the AI's working memory, "increasing response time and costs". Ten connections you never use make every answer slower.
Seven checks before you connect anything
- Use the vendor's official server. Find it on the vendor's own site, not through a search for a download.
- Connect the fewest tools that do the job. Three you use beat fifteen you might.
- Read the permission screen. Prefer read-only access when it is offered.
- Keep approval switched on for anything that sends, pays, deletes or publishes.
- Do not mix sensitive and open. An assistant connected to your accounts and also reading arbitrary web pages is the risky combination.
- Review connections every quarter and remove what you stopped using.
- Write it into your rules. Our AI policy template has a place for which tools staff may connect.
How an AI agent uses connections for you
ChatGPT, Claude, Gemini, Microsoft Copilot, Manus and Meta's Muse all use connectors of this kind to reach your tools. The practical question is who does the wiring and who watches the result.
- Describe the outcome. On Jobbit you say what you want, such as a weekly sales summary in your inbox, and the agent builds it.
- A plan before action. Plan mode lists the steps and waits for your approval.
- Secrets kept apart. Projects store keys and instructions separately from the conversation.
- Hosted for you. The automation or app runs on Jobbit, on your own domain from £7 a month on Go. See the plans.
- People when needed. A vetted professional from the human network can set up or review anything sensitive.
For ideas, see 50 AI automation ideas for small businesses, and for the basics, how AI agents work.
Frequently asked questions
What is MCP in simple terms?
MCP is a standard way for AI assistants to connect to other software. It works like a universal plug: a company builds one connection for its product and any assistant that supports the standard can use it.
What does MCP stand for?
Model Context Protocol. "Model" is the AI model, "context" is the information and tools it is given, and "protocol" means an agreed set of rules for how the two sides talk.
Who created MCP and who owns it now?
Anthropic created it and announced it on 25 November 2024. In December 2025 it was donated to the Agentic AI Foundation, part of the Linux Foundation, whose top-tier members include OpenAI, Google, Microsoft and Amazon Web Services.
Does ChatGPT support MCP?
Yes. OpenAI announced support in March 2025 and built its apps inside ChatGPT on the protocol in October 2025. Which plans can add custom connections has changed several times, so check OpenAI's current help page.
Is MCP safe to use?
It is as safe as the servers you connect and the permissions you grant. Use official servers, connect only what you need, keep approval on for actions that send or spend, and avoid combining access to private data with reading untrusted content.
Do I need to understand MCP to automate my business?
No. It helps to know the idea, but you can describe the result you want and let an agent handle the connections. On Jobbit the agent builds and hosts the automation and shows its plan first. Try it free with 100 credits a day and no card.