MCP, without the hype layer

Model Context Protocol — an open standard for connecting AI applications to tools and data. Anthropic introduced it in late 2024, and the industry has adopted it broadly since. One protocol replaces N bespoke integrations: a client (your AI app) speaks MCP to any server, which is a wrapper exposing a database, API, or capability. The model can then discover and call those tools. The official spec and docs live at modelcontextprotocol.io, which outranks this or any explainer for the current details.

Tool calling is the model-side mechanism. MCP standardizes the plumbing around it: how tools are described, discovered, transported (locally or remotely), and shared across applications. Before MCP, every app hand-wired its integrations; after, a server written once works in any MCP client. For teams the practical consequence is that integration effort moves from per-app glue to per-capability servers. It also moves to vetting the servers you adopt, which our evaluation guide treats as the real new skill.

The protocol is neutral. The risk concentrates in which servers you run and what you let them reach. An MCP server executes with the permissions you give it, and prompt-injected models can be steered into misusing powerful tools. So the evaluating-servers page is this site's core: provenance, permission scope, and transport security. It also covers the injection question every deployment should ask before wiring a server to anything that writes.