Notes
🔍

MCP Client

An agent that needs your files, a database and a wiki once needed three bespoke integrations. The Model Context Protocol replaces them with one standard connection, and the MCP client is the half of that connection living inside the agent application — the component that speaks the protocol to a server on the agent's behalf.

What the client is responsible for
One client holds one session with one server, and a host application runs as many clients as it has servers connected. Its work:

  • Connect over a transport — a local process over stdio, or HTTP for a remote server — and negotiate protocol version and capabilities;
  • Discover what the server offers: its tools, its readable resources, its prepared prompts;
  • Translate — present those tools to the model in its own function-calling format, then turn the model's chosen call into a protocol request;
  • Return results, errors and progress into the conversation, and surface anything needing user approval.

Desktop assistants, IDE extensions and SDKs such as the Claude Agent SDK are all MCP hosts in this sense: the client is plumbing they contain, not something the user configures beyond naming which servers to connect.

Why it matters
The client is where interoperability becomes real. Because it implements a published protocol rather than one vendor's API, any server the community writes works with any client, and an integration problem that used to multiply collapses into servers on one side and clients on the other. It is also the natural place to enforce trust: every tool call passes through the client, so scoping, logging and approval gates belong there.

An MCP client is the component inside an agent application that holds one session with one MCP server: it connects over a transport, discovers the server's tools, resources and prompts, converts the model's chosen call into a protocol request, and returns the result to the conversation. It is the half of the Model Context Protocol the agent owns, and the point at which any server works with any host.

Think of a standard wall socket. Appliances made by anyone fit it, because both sides agreed on the shape of the plug.

An MCP client is the socket built into AI assistant software. When the assistant needs to reach something outside itself — your files, a calendar, a company database — the client is the part that makes the connection, asks the other side what it can do, passes on the assistant's request, and brings back the answer.

Because the shape of the plug is published and shared, anyone can build a service and every assistant with one of these sockets can use it straight away. No custom wiring for each pair.