What is the Model Context Protocol (MCP)?

Before MCP, every connection between an AI assistant and an external system was a custom integration, built once per assistant per tool and maintained forever. MCP standardizes the plumbing, which is why it is often described as a USB-C port for AI: build one MCP server for your service, and any MCP-compatible assistant can use it. Anthropic open sourced the protocol in November 2024, and adoption spread across the industry through 2025, including among competing AI labs, which is what turned it from a vendor feature into infrastructure.

The architecture is client-server. The assistant runs an MCP client; a service exposes an MCP server that declares its capabilities, tools the model can call, resources it can read, and prompts it can use, each with a name, a description, and a typed schema. The model reads those declarations, decides what to call for the task at hand, and gets the results back as context it can reason over. Servers can run locally alongside the assistant or remotely over HTTP.

Two nuances matter in practice. First, a server is only as useful as its descriptions, because the model chooses tools by reading them; vague descriptions produce an agent that guesses. Second, wiring assistants to live systems raises genuine security questions, around what a tool is permitted to do and around untrusted content flowing back through it, so scoping, auth, and review of what a server exposes are part of the job, not an afterthought.

How this shows up in Waldo

Waldo ships a native MCP server exposing the same 200+ tools as its API, with the same auth and the same response shapes, so the brand intelligence an analyst reaches by code is available to an agent in conversation: brand_overview, discover_ads, audience_insights, category_trends, and the rest. It is in open beta, with a self-serve key at /build/get-started.

Related terms and reading