MCP explained simply: how an AI connects to your tools
The Model Context Protocol in plain words: servers, tools, OAuth sign-in and safety annotations, plus the questions to ask before connecting an AI.
The Model Context Protocol (MCP) is an open protocol introduced by Anthropic in late 2024. It defines a common way for an AI assistant to use external tools: read a document, query a service or, with SteerLatch, act on a computer.
Three ideas to remember
- The client: the assistant you use (Claude, ChatGPT, Grok…).
- The MCP server: the service that offers tools. SteerLatch’s is at
https://app.steerlatch.com/mcp. - Tools: the available actions, each described with a name, parameters and annotations (read-only, destructive…).
How the connection works
For a remote server, the MCP specification relies on OAuth: you sign in with the server’s provider and explicitly authorise the assistant. No password is given to the AI, and the authorisation can be revoked.
Before connecting an AI to your tools
- Which tools are exposed, and which can modify your data?
- Which devices or accounts do they act on?
- How do you revoke access and stop an action in progress?
- What data goes to the AI provider?
A shared protocol does not by itself make an application safe: check the permissions of the server you connect.