Glossary
MCP (Model Context Protocol)
The Model Context Protocol is an open standard through which AI applications access external tools and data sources in a uniform way.
On this page
MCP was published by Anthropic in 2024. The idea: instead of building a separate connection for every combination of AI application and software, a piece of software provides an MCP server once. Every AI application that speaks the protocol can use its functions, for example looking up a contact or creating a ticket.
An MCP server describes its tools with name, purpose and expected inputs. The language model decides during the conversation which tool it calls. For phone assistants this is interesting because systems can be connected for which no ready-made integration exists. Permissions remain important: a connected tool can read and change data. Which actions are allowed and what has to be confirmed before execution should be laid down.
In practice
You connect an MCP server by entering its address and credentials in the AI application. After that you see the list of tools it offers. Enable only those the assistant needs for its task, and use an account with restricted rights. On the phone, speed matters as well: every tool call lengthens the pause in the conversation. So check how fast the server responds and consider which queries really have to run during the call and which can wait until afterwards.
An example
A staffing agency works with an applicant tracking system for which there is no ready-made connection to its phone assistant. The software’s vendor does provide an MCP server, however. Through it the assistant gets two tools: “find application by name” and “add note to application”. When an applicant calls and asks to be called back, the assistant finds the record and leaves the request there. A third tool, which changes the status of an application, stays switched off on purpose.
A common misunderstanding
MCP is sometimes misread as a security feature. The protocol defines how tools are described and called, not what the assistant may do with them. You draw that line yourself through permissions. Nor does MCP replace a well-thought-out connection: twenty enabled tools do not make an assistant more capable but less predictable, because it picks the wrong one more often.
Source: modelcontextprotocol.io
Related terms:
On this topic:
As of: 12 October 2026