Dictate into whatever window you were working in, fix the text by talking to it, then paste. Or hand the draft to the agent CLI you already have, work the prompt over, and paste the version you settled on.
Speech recognition runs on your machine. No API key.
English only · three dependencies · Windows in full, macOS and Linux in Lite
Recorded by scripts/reel.py against the real windows — no
microphone, no model, nothing pasted.
uv tool install git+https://github.com/samartomar/flow
flowNo Python? Take
flow-windows-x64.zip
instead — that link always serves the newest release. It is unsigned, so the first launch
shows Windows SmartScreen and takes More info → Run anyway, once.
You need uv, which fetches Python 3.12 and the three
dependencies itself, and a microphone.
An agent CLI — codex or claude on PATH, already signed in — is optional. It adds
semantic rewrites, the prompt polish and converse mode. Without one, startup prints
refine CLI: NONE and everything else works.
The guide has the rest: cloning to change it, trimming ~106 MB
of unreachable dependencies, and why a .cmd launcher from npm -g is refused.
Off Windows, Flow starts in Lite without being asked — --lite runs the same code on
Windows if you want to see it. Startup says which one you got:
Flow Lite on darwin: Send copies the draft and you paste it - no injection,
no global hotkeys, nothing to grant but the microphone.
Lite is the brain, the ear and the clipboard. Both decoder tiers, the correction loop, the lexicon, the calibrated profile, the prompt polish, the thread and converse mode are all there, unchanged. What changes is the last inch: Send copies the draft and you press Ctrl+V yourself.
Four things it does not do — exclusions, not gaps: no injection into another application's window, no global hotkeys (arming is a click on the pill), no auto-paste, and no target-window awareness. What that buys is the property full Flow cannot have: nothing to grant but the microphone — no accessibility permission, no input monitoring, no trusted-application prompt.
Two requirements Lite cannot meet, named rather than dropped: P7 (safe paste into a terminal) is a promise about a paste Flow performs, and Lite performs none; and P9's loop ends on the clipboard, one keystroke short. docs/product.md defines the half and the fence around it — features land in full Flow first and reach Lite only if they survive without hands.
There is no Mac or Linux download. uv tool install is the way in, and a native macOS
body is not promised: it is weeks of work plus re-taking every measurement per OS, and
what would fund it is evidence from Lite.
- Click the pill (or
ctrl+alt+space) and talk. The draft floats up above it, refreshing about once a second. - Talk to the draft to fix it — "change Tuesday to Wednesday", "delete the last sentence", "scratch that". Anything that is not a correction is added to the draft instead. Or press Edit and type, which is faster for a URL or a flag.
- Send — the chip,
ctrl+alt+enter, or saying "boom" — pastes into the window you were working in.
Nothing sends itself. Stopping leaves the draft on screen waiting for you, and Flow presses Enter only when you ask for it by name ("enter boom").
Send does not erase, either: "bring back my last prompt" and "follow up: …" both reach what you already sent.
Dictate is the default: Send pastes.
Converse (ctrl+alt+M) turns Send into Ask. The draft goes to codex or
claude as a prompt to improve rather than a task to carry out, the answer renders in
the bubble and is read aloud, and Use this makes the answer your new draft. Point it
at a project with --cwd and the advice is about your code.
Developers who speak English as a second language, with a strong accent — Spanish, Indian, Russian and Japanese L1 speakers anchor the design and the benchmarks. That is what the correction grammar, the calibration pass and the personal lexicon are for. docs/product.md states why, and what it changes.
No API key is read, stored or passed anywhere in this codebase. Audio, the utterance buffers, the lexicon, the profile and every local edit stay put.
What leaves is what you hand to an agent CLI, which is cloud-backed: the draft tail on a rewrite, and the question plus the workshop preamble on an Ask. That preamble names your workspace, so a filesystem path leaves the machine along with the words — and a project path can identify an employer or a client. Set no workspace and nothing of the kind is sent.
Send also puts the draft on the Windows clipboard, where any clipboard manager or cloud-clipboard sync you run will see it.
One optional extra opens a socket: the [edge] voice pack sends the text of each spoken
reply to Microsoft to be synthesised. Install neither extra, or pick any other voice,
and nothing is sent. docs/architecture.md § What leaves the
machine states the boundary precisely.
The full list is in the guide. The four worth knowing before you start:
- Corrections have to be phrased as commands. "delete the bit about the standup" works; "I feel it should not contain the summary" is appended to your draft as text. The first recording from an Indian-L1 speaker phrased every correction the second way and 0 of 10 were recognised. The cause is register rather than accent, and it is unfixed — Refine forces the next utterance to be read as an instruction, and Edit lets you type the fix instead.
- A personal lexicon cuts both ways. Measured on EdAcc with
small.en, it recovers 27–34% of the rare words the model otherwise missed, and makes WER 14–38% relatively worse on speech containing none of the terms. Add words you say often, not every word you know. - Partials refresh about once a second, not per word, and can contain nonsense mid-word. They are shown dimmed and always replaced by the final text.
- Accuracy on your own voice is unmeasured. The per-accent numbers come from
recordings of other people.
scripts/live_check.pymeasures yours.
| docs/guide.md | the manual — every flag, every spoken command, both modes, voices, calibration, vocabulary, what is stored where |
| docs/product.md | what Flow is for: the target user, the P1–P9 requirements, the non-goals |
| docs/architecture.md | runtime reference: data flow, threads, the event stream, the tuning constants and their measurements |
| docs/roadmap.md | the gap between the product definition and this build, with the accent benchmark that measures it |
| docs/analysis.md | requirements breakdown, the four architectures considered, ranked risks |
| docs/development.md | layout, tests, the harnesses, the scripts, building a distributable |
| docs/history/PROGRESS.md | the build log, including the things that broke |
Collaborators welcome — especially anyone who speaks accented English and can tell me where Flow mishears them. That is the one thing I cannot measure alone.
git clone https://github.com/samartomar/flow && cd flow
uv sync && uv run flow # run it
uv run python -m unittest discover -s tests # 1,881 tests, ~40 s, no mic needed
uv run python scripts/selfdrive.py # the end-to-end harnessOpen an issue for a mishearing — what you said, what appeared — or a pull request for anything else. docs/development.md has the layout, the harnesses and the scripts; docs/decisions.md is why things are the way they are.
MIT licensed.