Model Context Protocol (MCP)
Model Context Protocol (MCP) is a standard interface for AI applications to discover and invoke tools or retrieve context exposed by connected servers. It gives an application a consistent way to learn what a server offers instead of requiring a custom integration shape for every tool.
How MCP works
An MCP client connects to a server and asks what capabilities are available. A tool definition describes its name, purpose, and expected inputs. The client can present eligible tools to an AI agent, send a structured invocation when the agent selects one, and return the server’s result to the conversation.
The protocol standardizes this exchange, but it does not decide which tools an agent should receive or when a tool call is safe. The application still controls server access, credentials, tool selection, timeouts, and how results are shown to the agent. A server also remains responsible for validating inputs and enforcing permissions.
Transport matters because the client must be able to reach the server. Some MCP servers run as local processes, while others are available over a network transport. Authentication and secret handling should match that deployment. Tool descriptions should be narrow and clear so the agent can distinguish similar actions.
Why it matters for AI phone calls
An AI phone agent often needs to do more than speak. During a call, it may need to look up an approved record, check availability, or create a follow-up item. MCP can expose those operations in a common tool format, reducing the integration-specific logic in the agent layer.
The live call still needs explicit failure behavior. A tool can time out, reject an input, or return an ambiguous result. The agent should avoid claiming success until the operation is confirmed, ask for clarification when required fields are missing, and hand off when the action is sensitive or cannot be completed safely.
In practice on ThunderPhone
ThunderPhone can add remote MCP servers by URL, synchronize them for automatic tool discovery, and attach their tools to agents. Static HTTP headers are supported for server access. The supported transport is Streamable HTTP; local servers that provide only standard input and output transport are not supported.