Model Context Protocol (MCP)

Before MCP, every AI application integrated every tool bespoke — an M×N problem. MCP standardizes the interface: an MCP server exposes tools, resources, and prompt templates from some system (a database, GitHub, a browser, an internal API), and any MCP-capable client can use them without custom glue. Servers and clients speak a common JSON-RPC-based protocol locally or over the network.

Anthropic released MCP as an open standard in November 2024, and it crossed the adoption threshold in 2025 when OpenAI, Google, Microsoft, and the major developer tools added support, making it the de facto industry standard for agent-tool connectivity. Thousands of community and vendor servers now cover common SaaS products, databases, and infrastructure.

For teams building agents, the economics are about integration cost rather than tokens: a connector built once works across models and applications, and switching model vendors no longer strands the tool layer. Attached MCP tool definitions do consume input-token budget like any tool schema, so loading only the servers a task needs keeps both context and cost tight.