Where is MCP's client? You never wrote it — but you're using it

Almost everyone learning MCP hits the same confusion:

"Aren't you supposed to have a Client? But I never wrote one, and opencode / Cursor just connect to Playwright and it works — where's the client?"

Here's the answer up front:

The client is always there. It's tucked inside the Host — transparent to you, which is why you never "feel" it.

Who's who in the "trio"

MCP has three roles — Host / Client / Server — and people immediately ask: "How many do I write?"

Role	What it is	Who writes it
Host	your LLM app (opencode / Cursor / Claude Desktop / your own agent)	the framework
Client	a piece inside the Host that connects the server, handshakes, calls methods	built into the framework (you don't)
Server	the process exposing tools	you write it

Key point: the Client is not a standalone program — it's part of the Host.

If you didn't write the client, who's doing the work?

You write one line of config:

"mcp": { "playwright": { "command": ["node", ".../cli.js"] } }

That just tells the Host: "launch / connect this server." After that:

[Client built into the Host]  →  initialize handshake
                              →  tools/list (fetch tools)
                              →  tools/call (invoke tool)
                              →  feed the result back to the model

It's all the Host's built-in client. You've been using a client all along — it's just transparent, because it lives in the framework.

An analogy

A browser connects to websites without you ever writing a TCP/IP stack — because the stack is built into the browser.

MCP's Client is to a Host what the TCP stack is to a browser.

What you configure is "which site to connect to" (the server), not "how to write TCP" (the client).

The right way to read the "trio"

• Host (framework) = built-in client + config-driven;
• Server = the tool you write;
• Client = the framework's, invisible, and not yours to write.

You really only need two things: write the server + add one config line.

So when do you write a client?

When the Host is your own program.

E.g., if you're building an agent from scratch — then there's no ready-made client, and you have to put one in yourself (or use an MCP client library). That's exactly why writing a standard MCP client by hand exists.

Remember one line

MCP always has a client. The only question is: does the Host ship it, or do you write it?

Most people get scared by the "trio" and think they must write three things. In reality:

You write the server (+ one config line). The client is already done for you.