What is MCP, and does your phone system need one?

What is MCP?

A common way for AI to plug into the software your business already uses.

An AI assistant on its own only knows what it was trained on and what you tell it. To check your diary, look up a customer or see who is waiting in a call queue, it needs a connection to the system that holds that information. Until recently, every one of those connections had to be built separately for every AI product.

The Model Context Protocol describes itself as “an open-source standard for connecting AI applications to external systems”, and compares itself to a USB-C port for AI. Build one MCP connection for a system, and any AI tool that speaks MCP can use it.

Anthropic launched MCP in November 2024. In December 2025 it donated it to the Agentic AI Foundation, a fund under the Linux Foundation co-founded with Block and OpenAI, with Google, Microsoft and AWS among its supporters. By then ChatGPT, Gemini and Microsoft Copilot all supported it. In short, it is no longer one company’s format: it is becoming the default way AI connects to business software.

Two words are worth knowing. An MCP server is the connector a system offers: it lists what the AI is allowed to do there, such as “look up a customer” or “book a meeting”. An MCP client is the AI product that uses those connectors. You rarely see either; your provider sets them up.

What does MCP have to do with a phone system?

It works in two directions, and both already exist on phone systems sold in South Africa, though one is still in beta.

An AI on your calls reaching out to your other systems. An AI receptionist that can only talk is limited. One that can check an order, find the caller in your CRM or book an appointment while the caller is still on the line can finish the job. MCP is one way it gets that reach.

An AI assistant reaching into your phone system. The other direction lets a manager ask an assistant such as ChatGPT or Claude about the phone system itself: who is waiting in the sales queue, where is yesterday’s call with a client, how did support do this week.

Yeastar lets its AI Receptionist use MCP servers during live calls, with examples such as tracking an order, identifying a caller from the CRM, booking into a shared calendar and opening a support ticket. A phone system can connect to up to 20 MCP servers, and each AI Receptionist can use up to five. More in our Yeastar AI Receptionist teardown.

3CX works the other way round. Its MCP Server lets AI clients such as ChatGPT and Claude view active calls and queues, search recordings and voicemails, generate reports, and, with permission, log agents in and out of queues or manage live calls. Access follows the connected 3CX user’s own permissions, and database queries are read-only. It arrives with 3CX V20 Update 10, which was still in beta at the end of September 2026. See our 3CX review.

For teams building their own agent, ElevenLabs agents can connect to external MCP servers. It is switched off by default, and ElevenLabs recommends having the agent ask for permission before each tool use.

Do you need MCP for your phone system?

Probably not directly. It depends on what you want the AI to do.

MCP is something your providers build and switch on, not something you buy. Whether it matters to you comes down to one question: does the AI need to reach a system it is not already connected to?

If you are in the first two rows, what matters is the list of systems the AI already works with. If you are in the last two, MCP support is a quick test of whether a phone system or AI provider is built to connect. It is a newer version of the question we raised in how to choose a future-ready phone system, where only three of ten phone systems we track published a documented developer API.

What should you ask your phone or AI provider?

Five questions that tell you more than “do you support MCP?”

What are the risks of connecting AI to your systems?

An AI can do whatever its connections allow, so the connections are the thing to get right.

The vendors themselves are cautious. ElevenLabs states that customers are responsible for the security of any third-party MCP server they connect, and switches the feature off by default. 3CX limits the AI to the connected user’s own permissions and keeps database queries read-only. Yeastar connects with fixed API keys or tokens rather than individual logins, which makes it important to keep those keys private and change them when needed.

None of this is a reason to avoid connected AI. It is the same discipline you would apply to a new staff member’s access, applied to software. For what goes wrong in practice, see why AI voice projects fail in South Africa.

Keep reading