We4AI Lab: Agent #1 — MCP Tool-Use Agent

What it does

MCP Tool-Use Agent connects a language model to external tools through the Model Context Protocol. The agent runs as a host application, opens MCP client connections to one or more MCP servers, and calls the tools those servers expose — the model decides which tool to call and the protocol carries the request and result. The full build is walked through in the video above.

Why I built it

Tool use is where agents stop being demos and start being systems, and MCP is the cleanest separation of “agent” from “tools” available right now. I wanted to learn the protocol by building with it end to end — host, client, and server — rather than reading about it.

Architecture

The build follows MCP’s three-part split:

user ──► host (the agent app + LLM) ──► MCP client ──► MCP server ──► tool
                                           │
                                           └────► MCP server ──► another tool

The host owns the conversation and the model. Each MCP client holds one connection to one MCP server. Each server advertises its tools (names, descriptions, input schemas) and executes calls. Adding a new tool means starting a new server — the host code does not change.

What worked, and what broke

A detailed failure log will be published with the code and follow-up posts. The honest summary so far: the protocol separation itself is the win — tools plug in cleanly once the host–client–server plumbing is in place — and the plumbing is where the time goes.

Cost, latency, and eval numbers

Not published yet. A measurement harness (fixed task set, per-run cost and latency logging) is the next piece of this project, and the numbers will appear on this page with their method and date — this site does not publish unmeasured claims.

What’s next

• Publish the evaluation harness and its first results.
• Add a second MCP server to show multi-tool routing.
• Write up the host–client–server setup as a step-by-step post.