MCPs: What, Why, and How

Model Context Protocol (MCP) is a standardized interface designed to enhance the interaction between large language models (LLMs) and external tools, providing a machine-readable alternative to traditional APIs that are typically built for human use. MCP establishes a consistent framework that allows LLMs to discover and utilize external services dynamically at runtime, enabling them to execute more autonomous actions without requiring bespoke integrations for every tool. By sitting above APIs, MCP assumes structured inputs and outputs optimized for machine processing, with tool servers maintained by service providers who understand their APIs best. MCP enhances agent systems by facilitating dynamic tool discovery, separating responsibilities clearly between agents and tool servers, and supporting complex multi-step workflows. It is particularly beneficial for applications involving external, stateful tools not owned by the user, such as content workflows, developer tooling, internal operations, and autonomous agents. While it introduces challenges in observability due to its reliance on external servers, MCP turns tool calls into first-class events, allowing for more detailed tracking and analysis of the agent's decision-making processes. Ultimately, MCP fosters an ecosystem that enables interconnected systems for LLMs, offering composable capabilities akin to what APIs have done for software.