MCP (Model Context Protocol)

The de facto agent-to-tool protocol of 2026, with millions of downloads and every major provider on board. Four capability types, one promise: build a server once, any AI client uses it.

MCP (Model Context Protocol) is the de facto agent-to-tool protocol in 2026. It standardizes how an agent connects to external tools, data sources, and services. It was created by Anthropic but was donated to the Linux Foundation later on. MCP exposes four capability types: resources, which comprise read-only data; tools, which represent the executable actions; prompts, which are the reusable templates; and sampling, which describes the LLM completions. The idea is to create an MCP server once and let any AI client use it without modification.

Dynamic tool registration, an MCP capability, enables tools to change at runtime without agents restarting. New definitions propagate instantly to running sessions.

The protocol landscape also creates vulnerabilities. In one Endor Labs analysis of more than 2,600 MCP implementations, 82% used file system operations prone to path traversal. Security cannot rely on model-level instructions: least privilege per tool, input validation at the execution layer, sandbox isolation and filesystem path scoping remain required regardless of protocol.