# MCP: Make Your Application Usable by Agents

**Track:** Interoperability, Safety & Production — AI-Native Full-Stack Development — complete (30)
**Framework / surface:** web dev (Model Context Protocol)
**Level:** Advanced
**Prerequisites:** Tool Calling: Giving Models Capabilities, Identity, Sessions & Authorization

**In one line:** Expose application capabilities through a standard agent-facing interface instead of rebuilding proprietary connectors for every host.

## Theory, aesthetics & inspiration

Anthropic introduced the Model Context Protocol in November 2024 — "a USB-C port for AI applications," as its documentation puts it: one standard interface through which agents discover and use tools, resources, and context, replacing a combinatorial explosion of proprietary connectors. For web developers the significant shift is that a product now has two first-class surfaces: the human one rendered in a browser, and an agent-facing one described in machine-readable capabilities. The protocol has matured into web infrastructure — the 2026-07-28 revision of the specification made the core stateless and moved long-running work into extensions such as Tasks, with authorization aligned to modern OAuth practice — and, as of 2026, the major TypeScript agent toolkits consume MCP servers as ordinary tool sources. Building a small MCP server over capabilities you already shipped is a forcing function for good API design: names, descriptions, schemas, permissions, and error behavior that make sense with no visual UI in sight. The architectural principle underneath: never make the model scrape your interface when the application can expose the capability directly.

**Builder question:** If an external agent could not see your UI, how would it discover what your application can do and invoke those capabilities safely?