Claude and the Pi add‑on

The small moment that made this story interesting was not that Claude could shop on Amazon. It was that, in the middle of the task, it noticed a Raspberry Pi project needed audio hardware and stopped to ask a real build question: should it add a basic USB mic and speaker, or a ReSpeaker HAT instead (howtogeek.com). That happened in a hands-on test published by How-To Geek on April 7, 2026, after the writer gave Claude permission to control a Mac through Anthropic’s desktop app and browse Amazon for parts for a local voice assistant setup (howtogeek.com). The striking part was not the click automation. Browser macros have done that for years. The striking part was that the model recognized a fork in the hardware path and asked for clarification before continuing. That is exactly the kind of thing Anthropic has been trying to build into Claude’s “computer use” system. The feature, first announced in public beta in October 2024, lets Claude interact with software the way a person does: by looking at screenshots, moving a cursor, clicking, and typing, rather than calling a neat, structured API built for the task (anthropic.com). Anthropic described the core challenge in unusually concrete terms. Claude has to infer what is on screen, decide what matters, and then count pixels well enough to put the cursor in the right place (anthropic.com). That sounds mundane. It is not. It turns a chatbot into something closer to a general operator. The Amazon shopping test shows what that means in practice. The writer did not ask Claude for a shopping list. He asked it to find the parts needed for a Home Assistant voice assistant, check reviews, and add suitable items to the basket (howtogeek.com). Claude did not just search for “Raspberry Pi accessories” and dump random results into a cart. It paused at the audio step because audio hardware is where this kind of project branches. A cheap USB microphone and speaker is the easy route. A HAT is more integrated. The model treated that difference as a decision point, not as noise. The hardware choice it surfaced was also a sensible one. Seeed Studio’s ReSpeaker 2-Mics Pi HAT is a Raspberry Pi add-on board built for voice and AI projects. It includes two microphones, audio output options, and extra interfaces for expansion, and it is explicitly marketed for voice interaction and assistant-style builds (seeedstudio.com). In other words, Claude did not invent an exotic accessory. It picked out a real category of purpose-built hardware that makes more sense for an always-listening Pi assistant than a pile of generic peripherals. That matters because Raspberry Pi projects are full of these awkward, half-technical choices. The Pi ecosystem is built around HATs, add-on boards that sit on the 40-pin GPIO header and turn the base computer into something more specialized (raspberrypi.com). Some are for high-quality playback. Some add microphones. Some drive speakers. Some are meant for compact embedded builds rather than a desk full of dongles (raspberrypi.com). A human builder learns to see those branches almost automatically. In this case, the agent did too, and it did it inside a live shopping flow on a normal consumer website, one click at a time (howtogeek.com). Anthropic has warned from the start that computer use is still experimental and can be slow, clumsy, and error-prone (anthropic.com). The How-To Geek test said the same thing from the user side. Watching Claude work was “incredibly slow” (howtogeek.com). But slowness is almost beside the point here. The interesting thing is that the model was not merely following a script. It was navigating a messy interface, inferring what the project needed, and then stopping on one concrete question: the cheap USB mic and speaker, or the little board with two microphones and a 3.5 mm jack that sits right on top of the Pi (howtogeek.com, seeedstudio.com).