Kleoth 0.2.0 — hold a key, talk, get typed text

Kleoth 0.2.0 adds voice typing. Hold fn+shift anywhere on your Mac, say what you mean, let go, and a second or two later the text is in whatever app has the keyboard focus — the Claude prompt box, a Slack message, a commit message in the terminal. Not a transcript of what you said: the text you would have typed. It’s opt-in, it uses your own ElevenLabs and OpenRouter keys, and it’s in the DMG at github.com/ofcRS/kleoth/releases/tag/v0.2.0.

Why build this into a meeting recorder

I had been using Wispr Flow for a few weeks and liked the shape of it — a key you hold, a pill at the bottom of the screen, text that appears. Two things pushed me out. The free tier is 2,000 words a week, which I burn through by Tuesday, and the paid tier is $15 a month. And it handles my speech badly: half of what I say is Russian with English technical terms in it, and whatever model they run kept flattening that to one language or the other.

Kleoth already had the pieces. It talks to ElevenLabs Scribe for cloud transcription of meetings, and to OpenRouter for summaries, with keys stored in the Keychain. Scribe v2 is the best speech model I’ve used for code-switching, and it bills per second of audio. A twenty-second utterance costs about a tenth of a cent there, plus a smaller fraction for the editing call. My first morning of heavy use came to a few cents. I pay for what I use, and nothing in the loop is somebody else’s subscription.

What happens between key-down and paste

The chord is watched with global NSEvent monitors, which is what the Accessibility permission is for. A pure state machine decides what a press means: a tap under 300 ms is ignored (that’s you reaching for an arrow key), a hold is push-to-talk, and a double-tap latches hands-free until you tap again. The microphone is opened at key-down, before the pill even appears, so the first syllable isn’t lost.

On release the clip is mixed to mono, loudness-normalized, encoded at 64 kbps, and sent to Scribe v2 with no_verbatim (it drops fillers server-side) and up to a hundred terms from a personal dictionary as keyterms — the names and jargon it would otherwise mishear. The raw transcript then goes through one OpenRouter call with a strict JSON schema, on google/gemini-3.5-flash-lite by default, which comes back in about a second. Then a pasteboard write and a synthetic ⌘V, and your previous clipboard is put back half a second later. If the editing call is slow, blocked, or answers in the wrong language, the raw transcript is pasted instead within eight seconds, with an orange note on the pill. Text always lands.

Nothing is kept except text. Each day’s dictations go into ~/Kleoth/dictations/<day>.json and show up under a Dictations scope in the History window; the audio clip is deleted the moment the pipeline is done with it.

The editing depends on where you’re typing

This is the part I care about most. I’m not a native English speaker, and when I dictate a prompt for an AI assistant I’m thinking out loud: I restart sentences, I hunt for a word (“the approach, no, the strategy”), I remember the important thing last. A transcript of that is faithful and useless.

So the editing call is told what app the text is going into, and picks one of three modes:

compose — AI chats, editors, notes, documents, mail, browsers: reorder the points, merge the fragments, render “first… second… and also…” as a list, keep the word I settled on, drop the thinking-out-loud. Every substantive point is kept, a “maybe” stays a “maybe”, and nothing is invented.
chat — Slack, Telegram, Messages, Discord: fillers and punctuation only. Sentence order and voice untouched, because a restructured chat message reads like it came from someone else.
terminal — plain text, one line, and a spoken description of a command stays a description. It never turns “run the tests” into a command.

Here is a real one, said into Claude:

okay so um I want to I want you to look at the the recording controller because it’s it’s too big like it does everything um the the capture, the pipeline, the the list of meetings, everything. and I’m thinking maybe we should uh we should split it, no not split, extract, extract the the queue part, the pipeline queue, into its own, its own type. um also, wait, before that, actually first, we need tests for it because right now there is no tests for the app package at all […] don’t do it yet, just just propose the plan, like how you would do it, and what what risks you see.

And what got pasted:

Look at the recording controller because it is too big and does everything: the capture, the pipeline, the list of meetings, and more.

Add tests first, because right now there are no tests for the app package at all, so if we change something we won’t know if we broke it.

Extract the queue part, the pipeline queue, into its own type.

Move the meeting errors dictionary together with the queue because it shares the same responsibility.

Don’t do it yet — just propose the plan and what risks you see.

The same sentence said into Slack comes out as one casual line. Russian stays Russian, and a Russian sentence with pipeline queue and MainActor in it stays exactly that mixed. I benchmarked six of these rambles across eight models before settling on the default; the lite Gemini was the fastest and, to my surprise, the most faithful. Claude Haiku 4.5 was as good and twice as slow. A couple of models timed out or truncated under the strict schema and are not offered.

The pill

Between dictations a small dark tab rests half off the bottom edge of the screen. Hold the chord and it springs out into a bar with a live waveform; release and the bars carry a wave while it transcribes and edits, then a check, then it sinks back. No words in any of that — words only appear for a warning or an error, because those need a reason and sometimes a button.

You can drag it along its edge, or drag it clearly toward another edge and it docks there; on a side edge it stands up and everything, waveform included, runs vertically. The animation itself was the hardest bit of UI in the release. Animating an NSWindow’s frame with AppKit’s animator while SwiftUI animates the content inside it gives you two clocks and two curves fighting, and it looked like it. The window is never animated now: on each move it is set, instantly, to a transparent stage covering both the start and the end rect, one SwiftUI spring carries the capsule’s offset, size and rotation, and when the spring settles the window shrinks back to the capsule. A second spring, staggered by ninety milliseconds, handles the shape change, with a quick squash-and-stretch as anticipation. It reads as one elastic motion.

Also in 0.2.0

Transcribing a meeting after recording is now opt-in. Stopping saves the audio and lists it as Untranscribed; you pick the engine per meeting, on-device or cloud, when you want it.
Per-engine transcript variants. Re-transcribing with the other engine keeps the previous transcript and summary; switch between them from the tier badge.
Remove Transcription reverts a meeting to its audio (transcript and summary go to the Trash), and rows now show their size on disk.
Playback plays both sides in both ears — the two-channel file is downmixed live instead of giving you yourself in the left ear.
Failures are visible where you’re looking: an error card on the meeting and a red chip on its row, instead of a line in the menu-bar popover you’d never see.
The Slack integration is gone; Copy Summary replaces it.

Rough edges

Dictation needs the Accessibility permission, for the hotkey and for the paste. Kleoth is deliberately not sandboxed — the sandbox blocks posting the synthetic ⌘V outright, so there is no Mac App Store path for this feature.
The paste is a synthetic ⌘V on the V key’s QWERTY position. Russian and other QWERTY-family layouts work; Dvorak and Colemak are not handled yet.
It is not free: a few tenths of a cent per utterance across two providers, on your keys.
Still self-signed, still right-click → Open on first launch.

Download: github.com/ofcRS/kleoth/releases/tag/v0.2.0. Full notes in the changelog.