Kettle Cups
The small jobs you might give an AI chatbot, done on your own computer.
Kettle is a collection of small tools - I call them cups. Some run small AI models locally, with nothing sent to the cloud. Some need no AI at all, only ordinary code.
Each tool is a Cup: it does one job, gives you a result you can check, and lets you save it. Each is published with a check of what it replaces and where it fails.
In development · not available to download yet. Always check the result before using it.
Try one in your browser
Geo, Regulars, Tables and Text each have a demo: the tool’s own page from the app, on samples I prepared. Nothing of yours is uploaded. A demo needs JavaScript.
Tools
Some run small, specialist AI models on your computer. Some need only ordinary code. Every tool page names the models it uses, or says if it uses none.
- Tables turns a photo of a printed table into a CSV, using four small models of about 34 MB in all, and reads PDF bank statements with no model.
- Meeting Notes transcribes a recorded meeting with a speech model on your computer, for you to listen back, correct and save.
- Text sorts and tidies text with ordinary code. No model at all.
- Share Copy takes the personal details out of a letter or email before you paste it into a chat model.
- Regulars lists what you pay for again and again from a year of statements, with ordinary code.
- Local finds a passage, a picture or a video in a folder of your own files, matching by meaning with a model on your computer.
- Media makes a smaller copy of a video or a picture, crops a picture or trims a recording with ordinary code. No model at all.
- Geo finds the photos in a folder by where they were taken, from the position recorded in each file, with ordinary code.
- Sort suggests where a pile of unsorted photos belongs in the folders you already have, by date and by look, with Apple’s models built into macOS. It moves nothing until you run the script it writes.
Each one gives you a draft to check. Setup downloads any models once; after that, normal use sends nothing to the cloud. See what every tool page tells you.
Nothing the tools do is particularly complex and that’s the point. People are increasingly using chatbots for all kinds of things, including the everyday mundane tasks that were previously done manually or by software. The question Kettle wants to ask is why Frontier AI should be used to order a list?
None of the Cups acts for you: none sends, books, files or calls. Each works on your own files and stops there for you to check.
Local or remote
For a repeatable task, what is the smallest thing that does it well enough: ordinary code, a small specialist model, a local language model or a remote model? Any of them can be the right answer, and so can what your computer already does.
Useful local models are often not language models. On one tested computer, an Apple M1 Pro, the Tables reader took about a second for each of seven small tables and used about 550 MB of memory at its peak.
Local means everything stays on your computer. No round trips to a remote data center using large amounts of resources to re-format a document.
Slow by design
Local means it uses your own computer's processing capabilities - that might mean it's slower that the latest chip from a trillion dollar company but I think that's no bad thing. Start a process on Kettle whilst you put the kettle on.
Research
During the development of the Cups I've tried out various things - sometimes it's with the idea that a local model will provide better results than deterministic code, other times it's just experimentation. More often than not I am wrong. I feel it's important to openly show the working behind these experiments, this is the section where I do that.
The first study tested my own idea - I gave a frontier agent a guide to Tables, expecting it to use less model work than reading the image itself. On seven small tables it used more. I published the result, the limits of the test and the token budget for one table that came out of it.
Read the first study, what has been checked about Tables, or all the research.
About
I’m Rich Holman, and I build Kettle Cups as DGW.ltd. I do not claim energy reductions, better models or reliability from a handful of examples. Read how I work and what I do not claim.