Receipt: your AI use, yours only

Counts your AI prompts, never what you type, and shows an honest energy and carbon estimate. Nothing leaves your device.

Receipt tells you what your own AI use costs, and is honest about what it cannot tell you. You have probably been told two things about AI and energy: that one prompt boils a lake, and that it is nothing, so stop worrying. Receipt does not settle that argument for you. It gives you your own number, as a range, with every factor sourced and graded. What it does - Counts the prompts you send on seven widely used AI chat websites. The full list is shown in the permission prompt when you install, and on the website linked below - Shows a weekly energy and carbon estimate, always as a low-to-high range - Flags when a prompt ran on a reasoning or "thinking" mode, which typically uses 20 to 60 times the energy of a standard prompt - Counts re-asks and regenerations, the one kind of waste entirely within your control - Lets you log what AI took off your plate, and what you did with the time - Includes a permanent panel on what your number does not cover: datacentre siting, water, chip manufacture, model training and the annotation workforce What it never does - It never reads, stores or transmits anything you type - It records three values per prompt: a timestamp, the site, and standard or thinking mode - It makes no network requests of any kind. There is no server and no account - It never compares you to anyone else, and has no scores, streaks or badges The useful finding For most people a small handful of reasoning-mode prompts account for the large majority of the energy, and most of those were drafting or summarising jobs a standard model does identically. That is free efficiency with no loss of quality, and it costs less time as well as less energy. Honest about its limits Receipt cannot see AI built into office software, meeting transcription, customer databases, design tools or mobile apps. Reasoning-mode detection relies on artefacts in the response and will occasionally get it wrong, so there is a manual override. Image generation is logged by hand. It shows no water figure at all. Published per-prompt water estimates differ by around 170 times because they measure different things, so any single number would mislead. The tool explains the absence instead. Energy factors come from the one provider measurement published with a full method, plus independent academic benchmarks. They are reviewed quarterly. Full methodology with every source linked: https://mariestorrar.co.uk/receipt/methodology Free, independent, and not selling you an AI product.