Mechanical Middleware
Mechanical middleware bypasses brittle corporate APIs by treating physical controls as a permanently backward-compatible interface.
By Theo Marsh
Sparked by Turning a dumb AC unit smart (without losing my security deposit) · discussion
Michael Prilik's blog post detailing his recent apartment automation project crossed my screen yesterday, and it looks completely ridiculous. He literally binder-clipped a cheap $2 stepper motor directly to the temperature dial of his window air conditioning unit. The Hacker News discussion immediately lit up with reactions to a rig cooked up with spare electronics and standard office supplies.
Your first instinct looking at that setup is probably the same as mine. Why bypass the internal electronics entirely? Why not pop the chassis, splice into the control board, or just buy a modern smart AC unit with a documented REST API? We are conditioned by decades of computing to view a motor physically cranking a plastic knob as a janky maker fever-dream. We assume it is a temporary stopgap waiting for Proper Architecture. That instinct is wrong. Prilik's rig is actually the most structurally sound software integration I've seen all year.
Let's call it Mechanical Middleware.
Contrast this approach with how we usually integrate appliances today. A proper software ecosystem requires betting the farm on a centralized cloud server that you do not control. When you hook into a corporate IoT network, you are attaching a ticking clock to your own hardware. We have seen the high-profile casualties repeatedly. Google summarily executing the 'Works with Nest' API bricked countless custom home integrations. The smart home company Insteon went a step further, shutting down its app servers overnight and abandoning thousands of households to dead silicon. Corporate integrations are massive architectural vulnerabilities dressed up as developer convenience.
To understand why Prilik's motor is superior, you have to look at the physical world as the ultimate DOM.
When web developers encounter a locked-down platform that refuses to offer an API, they rarely give up. They write screen-scrapers to read the HTML DOM directly. The structural elements of the page become the interface. If a div exists for a human to read, a scraper can read it too.
Appliance manufacturers are building the exact same walled gardens in our living rooms. Mechanical Middleware is how we bypass them. If the AC unit offers a physical knob for a human hand to grasp, it has exposed a public endpoint in the Physical DOM. Turning that knob with an external motor bypasses the vendor's cloud dependency entirely. The manufacturer cannot push an over-the-air firmware update to deprecate a piece of plastic.
The engineering benefits here are deeply pragmatic. Bypassing the electrical hazards of high-inductive AC loads to actuate a dial with low-voltage mechanics is simply the safest path for an integration. You avoid messing with mains power entirely. And while binder clips might look amateurish, the market has already quietly validated this exact architecture. Commercial products like the SwitchBot Bot are built explicitly to retrofit dumb appliances with smart, mechanical button-pushers. They package the mechanism in sleek white plastic, but the underlying principle remains identically robust. They are navigating the Physical DOM.
Widen the lens a bit, because this is the exact blueprint for how AI agents will operate in our physical environments.
Autonomous agents won't patiently wait for appliance manufacturers to document OAuth endpoints or agree on a universal interoperability standard. They will use Vision-Language Models to visually read analog dials & indicator lights -- effectively performing a physical GET request. We already know models like GPT-4V can interpret these visual environments reliably.
Once an agent has the state of the room, it will use cheap, disposable Mechanical Middleware to turn the dials and push the buttons. That actuation is the physical POST request. The agent doesn't care if the underlying appliance is smart or dumb, just as a scraper doesn't care if a website is generated by React or raw HTML. It only cares about the resulting state change.
A software API is a promise the vendor will eventually break. The physical world is backward-compatible forever.