In this guide · 8 sections
- 01What is the Model Context Protocol (MCP)?
- 02How MCP works: clients, servers and tools
- 03Which AI tools support MCP today
- 04Why MCP matters for a small business
- 05MCP security: questions to ask before you connect
- 06A sensible way to start with MCP
- 07Do it yourself or get help?
- 08Frequently asked questions
Short answer
MCP is an open standard for connecting AI assistants to apps and data, described by its maintainers as a "USB-C port for AI applications". Claude, ChatGPT and Copilot Studio support it. Before connecting, check what each connector can read and do, who approved it, and whether actions need confirmation.
What is the Model Context Protocol (MCP)?
MCP is a published set of rules for how an AI application talks to outside systems. The project's own introduction describes it as an open-source standard for connecting AI applications to data sources (such as files and databases), tools (such as search engines and calculators) and workflows, and compares it to a USB-C port: one standard plug instead of a different cable for every device.
Anthropic introduced MCP on 25 November 2024 as an open standard. On 9 December 2025, Anthropic donated MCP to the Agentic AI Foundation, a fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI. That matters for businesses: MCP is not one vendor's proprietary feature, so a connector built once can work with several AI tools.
The specification is still evolving. The current version at the time of writing is dated 28 July 2026, which made the protocol stateless and hardened authorisation, according to its changelog.
How MCP works: clients, servers and tools
You do not need to know the technical detail, but three words come up in every conversation with a vendor or developer.
| Term | What it is | Example |
|---|---|---|
| MCP client (host) | The AI app your staff use | Claude, ChatGPT, Copilot Studio, an AI coding tool |
| MCP server | A small program that exposes one system to the AI | A server for your CRM, Google Drive, Xero or a database |
| Tools | Actions the server lets the AI take | Search contacts, create invoice draft, update a ticket |
| Resources | Data the server lets the AI read | Files, records, reports |
A server can run on your own computer (local) or on the internet (remote, signed in with OAuth). In the AI app, these usually appear as connectors, apps or tools you switch on. When a staff member asks a question, the AI decides whether to call a tool, the server performs it with the permissions it was given, and the result goes back into the conversation.
Which AI tools support MCP today
Support is now widespread among the major AI platforms. From each vendor's own documentation, as of October 2026:
- Claude: custom connectors using remote MCP are available on Claude, Claude Desktop and related apps for Free, Pro, Max, Team and Enterprise plans, with Free users limited to one custom connector (Claude help). Team and Enterprise owners can add connectors for the whole organisation.
- ChatGPT: OpenAI's help centre explains connecting MCP servers to ChatGPT as apps, and its Apps SDK is built on MCP. Workspace admins control which apps are allowed.
- Microsoft Copilot Studio: Microsoft announced MCP is generally available in Copilot Studio in May 2025, so agents built there can call MCP servers.
- Developer tools: many AI coding assistants and agent frameworks also act as MCP clients.
If you already use one of these on a business plan, check its admin settings: connectors may be available to staff by default. Our guide to ChatGPT connectors for Google Drive and SharePoint covers the admin controls for the most common case.
Why MCP matters for a small business
Before MCP, connecting an AI assistant to your systems meant a custom integration for each tool, or copying and pasting data into a chat. MCP lowers that cost.
What it makes possible
- Ask an AI assistant about live data: open invoices, stock, bookings
- Draft replies or records directly in the system, for a human to approve
- Reuse one connector across Claude, ChatGPT and other clients
- Build a small internal connector instead of a full app
What to watch
- The AI acts with whatever permissions the connector has
- Third-party servers may be unvetted or abandoned
- Hidden instructions in data can manipulate the AI (prompt injection)
- Data leaves your system and goes to the AI provider as context
For simple, predictable tasks, such as sending a WhatsApp message when a form is submitted, a traditional automation is still cheaper and more reliable. Our Zapier vs Make comparison covers that route. MCP shines when people want to ask questions or take varied actions in plain language.
MCP security: questions to ask before you connect
The MCP project publishes its own security best practices, covering problems such as confused deputy attacks on OAuth proxy servers and token handling. You do not need to read them, but whoever sets up your connectors should. As a business owner, ask these questions.
Before connecting an MCP server
Singapore's CSA made similar points about autonomous agents in its 2026 advisory; see our guide to CSA's OpenClaw advisory.
A sensible way to start with MCP
Start small and read-only, then add actions once you trust the setup.
Pick one question staff ask every day
For example: which invoices are overdue, or what is a customer's last order. That defines the data needed.
Use an official connector if one exists
Prefer connectors published by the software vendor or listed in the AI platform's directory over unknown community servers.
Connect read-only first
Give the connector a dedicated account with read access to only the data needed for that question.
Turn on admin controls
On Team or Enterprise plans, restrict which connectors staff can enable and review the audit log.
Add write actions with approval
Once it works, allow drafts or updates that a person confirms before they are saved or sent.
If no connector exists for your system, a developer can build a small MCP server for it. That is often less work than a full integration, and it works with more than one AI tool. Our guide to what AI agents cost a small business gives realistic budgets.
Do it yourself or get help?
Switching on an official connector in Claude or ChatGPT for read-only access is a reasonable DIY step for a tech-comfortable owner. Building a custom MCP server, connecting it to business data with the right permissions and approval steps, and testing it against prompt injection is developer work.
Our Custom Tool or Dashboard service (from S$1,990) can build a scoped MCP server for your system. For fixed, rule-based tasks that do not need an AI in the loop, Workflow Automation (S$390 per workflow) is usually cheaper and more predictable.
Frequently asked questions
No. Anthropic created MCP but released it as an open standard and donated it to the Agentic AI Foundation under the Linux Foundation in December 2025. OpenAI documents MCP support in ChatGPT, Microsoft supports it in Copilot Studio, and many developer tools and agent frameworks act as MCP clients, so one server can work with several AI tools.
Not always. Many AI platforms offer ready-made connectors that you switch on and sign in to, with no code. You need a developer when there is no existing connector for your system, when you need tight permissions or approval steps, or when the connector will handle personal data or financial actions that must be tested properly.
It can be, with care. The protocol itself does not decide what the AI can access; the connector's permissions do. Use dedicated accounts with the least access needed, prefer read-only, require approval for actions, check the AI provider's data terms on your plan, and remember PDPA obligations still apply to any personal data the AI processes.
An API is how a specific piece of software lets other programs talk to it, and every API is different. MCP is a common layer on top: an MCP server usually calls a system's API and presents it to AI assistants in a standard way, so the AI can discover the available tools and use them without a custom integration for each assistant.
An MCP server is a small program that connects one system, such as a CRM, file store or database, to AI applications using the Model Context Protocol. It lists the tools and data it offers, and performs actions when the AI asks. It can run on your own computer or as a hosted service you sign in to.