Writing
Proof of Concept: Coding an MCP Client with Claude
I gave Claude a small, concrete target: build an MCP client that could pull NBA stats. Not because I needed the stats. I wanted to watch an AI coding agent work against a real integration boundary and see where it broke. The headline finding was not that Claude could write the code — it obviously could. The finding was how much of my time went into steering: scoping the task, verifying outputs, and pulling the implementation back on course when it drifted.
What MCP actually changes
MCP gives an assistant a clean way to reach tools and data. Once the agent can inspect an API, call a local server, and iterate on its own failures, the experience stops feeling like autocomplete. It feels closer to a junior engineer with absurd throughput — fast, tireless, and occasionally confidently wrong.
The POC forced the agent through the same gauntlet a real integration faces: discover the API shape, implement the protocol, handle configuration, expose a stable tool contract, and produce something another system can actually call. Those are exactly the failure points that show up in larger agent platforms, just in miniature.
Where the human still mattered
The agent moved fast, but it had no opinion about what "done" meant. I had to define the outcome, notice when the implementation was overfitting to the happy path, and decide which errors were worth fixing and which were noise. AI coding looks magical when the acceptance test is loose. It gets brittle the moment the acceptance test is real. The model is great inside a frame; the human still has to build the frame.
A few things I would do differently next time. Define the external tool contract before letting the agent touch internals. Give it a small set of real tests instead of prose requirements — prose is a suggestion, a failing test is a fact. And keep "working demo" and "production ready" as separate gates: auth, rate limits, logging, and error handling do not come free just because the demo runs.
So MCP is not magic. But it gives agents the handles they need to be builders instead of chat windows, and that turns out to be most of what was missing.