You said no MCP

MCP has moved beyond a simple code‑generation aid; the author reports that it now powers natural‑language configuration in macOS apps such as rcmd, Clop, and Lunar. With a local Qwen model and a Pi device, a user can say, “Set up Clop to optimise any PNG I drop in my website assets,” and the app adjusts itself accordingly. The claim is that this works reliably enough for everyday workflow.

The setup runs the language model locally, avoiding cloud latency and privacy concerns. Qwen processes the spoken or typed instruction, translates it into a series of CLI flags or API calls, and the Pi acts as a lightweight inference server. Because the model runs on the same machine, the round‑trip time stays in the low‑hundreds of milliseconds, making the interaction feel instantaneous.

The author notes that this approach mirrors other imperfect but widely adopted standards like USB‑C or HDMI: it isn’t perfect, but its broad compatibility wins out. MCP’s “sub‑agent” architecture lets a smart model delegate to smaller, task‑specific models, then reconvene for a final review. This pattern keeps the main model lightweight while still handling complex commands.

If you’re building a macOS utility, you can embed MCP to expose a natural‑language endpoint instead of a static preferences pane. Hook the model’s output into your existing command‑line interface, and you’ll give power users a way to script behavior on the fly without learning new syntax.

TakeawayEmbedding MCP lets desktop apps accept plain‑English commands, turning user intent into immediate configuration changes.