CrankGPT pitches a human-powered offline AI that runs on calories, not cloud

CrankGPT, a product site written with a straight face and a raised eyebrow, is pitching what it calls a human-powered, fully local and private AI solution. Instead of GPUs in leased data centres, its tiers are measured in watts of human effort: a basic hand-cranked model for everyday home use, pedal-powered models for power users and small companies, and partnerships with gyms and fitness studios for complex agentic and enterprise workflows.

Three tiers, measured in watts

The catalogue is deliberately familiar. Synapse, rated at 20W, handles question answering and simple conversation. Cortex, at 150W, is aimed at agentic coding and video generation. Singularity, at 2000W and above, is reserved for agent swarms, training and fine-tuning. The pitch is what the site calls rightsizing — use the appropriate tool for the job — and producing your own tokens by burning calories instead of oil.

Today's forecast: cloudless

The privacy argument is the serious part. The site asks why anyone would hand mega-corporations their most burning questions, innermost thoughts and half-formed app ideas, and says CrankGPT runs entirely on device so the data stays with its owner. It also lists what a cloud service cannot survive: no Wi-Fi, a provider outage, rolling blackouts. With the machine in front of you, the intelligence is still there.

Energy bills, supercars and tokenmaxxing

CrankGPT leans into the energy debate around AI, noting that technology companies have quietly abandoned climate pledges to build gas-burning power plants that feed their models; its counter-slogan is to stop burning oil and start burning calories. It also needles the industry's economics, arguing that users have been hooked on tokens they pay more for than they cost to generate, and that the money is better kept out of tech chief executives' pockets.

What is actually behind it

The site links to technical documentation and to a GitHub project describing a small voice agent that runs fully offline on edge devices, with a fitness-coach demo that echoes the exercise theme. That project supports interchangeable speech recognition (Moonshine, FasterWhisper, NeMo FastConformer, Vosk), text-to-speech (Piper, Kokoro) and local models served through llama.cpp, with a default setup that runs on a Raspberry Pi 5. Demo requests go to [email protected]. Whether CrankGPT ships hardware matters less than the inversion it performs: it takes the industry's real selling points — compute, scale, the cloud — and asks what happens when the compute is your own arms.