What Is MCP (Model Context Protocol)?

AI coding tools are increasingly capable, but they operate with a fundamental limitation: they can only reason about information available in their context window. When an AI agent reviews a pull request, it sees the diff — not the full dependency graph, not the blast radius, not the compliance implications. It infers structure from the code it can see, which is probabilistic and often incomplete.

MCP closes this gap by giving AI agents a structured way to query external data sources. Instead of inferring that "this file is probably important," an AI agent can query an MCP server and receive the file's PageRank score, SPOF status, and blast radius — deterministic data extracted from the full architectural analysis.

The protocol is transport-agnostic (supporting Streamable HTTP, SSE, and stdio) and tool-agnostic (any data source can expose tools). This makes MCP a universal integration point between AI agents and the data systems that provide their context.

MCP operates on a client-server model with four phases:

Initialization: The client (AI agent) connects to the MCP server and sends an initialize request. The server responds with its capabilities and supported protocol version.

Discovery: The client requests the list of available tools via tools/list. Each tool is described with a name, description, and JSON Schema for its input parameters.

Invocation: The client calls a specific tool by sending a tools/call request with the tool name and typed arguments. The server executes the tool and returns a structured response.

Session management: For Streamable HTTP transport, the server issues a session ID via the mcp-session-id header. The client includes this header in subsequent requests to maintain session state.

The protocol is stateless per-request but session-aware across requests, enabling servers to maintain context (like which analysis dataset to query) without requiring the client to re-authenticate on every call.