Kety
Source on GitHub
Today, you are expected to have an AI subscription for everything. One for dictation, one for meeting notes, one for rewriting, one for a chatbot that knows your work, each with its own account and its own monthly charge. They all need a copy of your work on their servers before they can do anything with it, and they stop working when the connection does.
None of that is required any more. Open source models small enough to run on a laptop do most of this work, which means it can be free, it can stay on your machine, and it can keep working with no connection at all. A recent Mac (or not) has what it takes to run them.
Kety puts that behind a few keyboard shortcuts. Anything on your screen can be kept with one: a paragraph you just copied, a screenshot, a recording, something you say out loud. It is saved and indexed on your machine as it arrives, so you can ask questions about it later, or have a model rewrite, translate or answer from it without leaving what you are doing. You install the models once, write the prompts, and decide which shortcut does what.
macOS blocks it on first launch, because it is not notarized with Apple. Open it once and you will be offered only Move to Trash; ignore that, go to System Settings › Privacy & Security, scroll to the bottom, and click Open Anyway next to the line about Kety. It opens normally from then on. On older macOS, Control-click the app and choose Open instead. Or skip all of that and build it yourself: an app you compiled is never put in quarantine, so it just opens.
Highlight text anywhere, in a mail draft, a document or a code comment, and press the shortcut. A small panel appears next to it with the actions you configured: translate, shorten, fix the tone, whatever you wrote a prompt for. Pick one and the rewritten text replaces your selection. If the answer is not quite right, keep talking to it in the same panel until it is.
Hold another shortcut and speak. Speech-to-text runs locally, then your own prompt rewrites the transcript into an email, into notes, into another language, and it is typed into the field you were in. Dictate over a selection and it becomes an instruction about that text instead.
The same shortcuts capture as well as transform. Whatever crosses your screen can be kept: a screenshot, a screen recording, a note you dictate, a PDF you drop in, the captions from a Google Meet call. It lands in a SQLite database in your own Library folder, and it is indexed as it arrives, so it is searchable a few seconds later without you doing anything.
The assistant answers from what you captured, and tells you which capture it read. The search runs on your machine against your own index, so a question about last week's meeting does not turn into a request to anyone's server.
Pick the captures you want to share and Kety builds a single file containing them and their index. Whoever opens it gets them as a read-only assistant alongside their own, and can ask it questions the same way. It is the fastest way to hand over what you know when you leave a project, or to give someone the context they would otherwise ask you for twice a day.
You can send a link for it, which is usually what people want, and the storage behind that link is yours. Point Kety at a Google Cloud Storage bucket you own and it uploads there and signs the link for you: your bucket, your credentials, your call on when it stops working. Nobody else is holding the file, and there is no account here that could be closed with your data inside it.
Or skip the link and send the file itself: by email, over AirDrop, on a USB stick. There is no service in the middle either way.
No paid tier, no trial, nothing to subscribe to, and no API bill, because the models run on your own machine, so the work costs you disk space and nothing else. Kety is released under the Apache 2.0 licence.
Capture, indexing, search and the assistant can all run on local models. Kety ships with no model weights, so you install the ones you want once, from Settings. After that it works with the network off: on a train, on a plane, in a basement meeting room, exactly the same as it does at your desk.
There is no Kety account, no sign-in, no password, and no Kety server to send anything to. There is no telemetry either: no analytics, no crash reporting, no update check, no startup ping. Your captures live in a SQLite database in your own Library folder, next to the media files. It is not encrypted at rest, so use FileVault if that matters to you.
The whole thing is open source. Read it, change it, fork it, or run a build of your own. What the app does with your data is written down in the code, which is the only claim on this page you do not have to take on trust.
Older Macs exist, and some jobs want a bigger model than yours can hold. You can plug in an OpenAI key and send those to a hosted model instead, chosen per task, so local stays the default and the key is there for what needs it.
It is a step away from the idea, and worth saying plainly. What you get back is one key instead of five subscriptions, billed for what you actually use. Rewriting a paragraph or cleaning up a dictation is a short request, so the cost per use stays small, and you are paying for the work rather than for the right to keep an app installed. The prompts, the shortcuts and the index stay yours either way.
The code lives at github.com/sverbo/kety, under the Apache 2.0 licence. The README there explains what the app does with your data in more detail, and is honest about the rough edges.
Issues and pull requests are welcome. Anything else, write to sam@kety.app.