Glossary
What is MCP?
The standard that lets agents talk to outside tools and data without a custom plugin per model.
Updated 2026-09-09
MCP is an open protocol for connecting an AI client to external tools and data through one shared interface, so the same integration works across editors and agents.
Before it existed, every tool integration was bespoke. A Figma plugin for one assistant, a different SDK for another, nothing reusable. MCP splits the problem in two. A server exposes capabilities: tools it can run, resources it can read, prompts it can offer. A client, meaning your editor or agent or desktop app, connects and uses them. Write the server once and every MCP capable client can use it.
Servers run either locally over stdio, which is how most filesystem, database, and CLI integrations work, or remotely over HTTP for hosted services. Anthropic published the protocol in late 2024 and the other large vendors adopted it through 2025, which is why a Postgres or Figma server you set up in one editor tends to work in the next one.
In day to day use, MCP is how an agent stops guessing. Instead of hallucinating your schema, it queries the database. Instead of describing a design, it reads the actual frame.
Where people get this wrong
- Every connected server injects its tool definitions into context. Ten servers you never use make the agent slower and less accurate, not more capable.
- A local server runs with your permissions. Read what it does before you point it at a repo or a production database.
- Tool descriptions are prompt surface. A vaguely described tool gets called at the wrong moment and you will blame the model.