Hold a key. Speak. Clean text lands where your cursor is.
Voice input for developers, on Windows. Speech recognition runs entirely on your machine — no account, no cloud, no telemetry.
winget install deixis soon
Free for everything that runs locally. Signed installer, ~10 MB; the speech model downloads on first run.
The part nothing else does
Annotate while you talk. Select text, grab the screen, point at elements — mid-sentence.
Tap Ctrl+Win+F and narrate. While you speak, capture what you mean: the text you have selected, a region of the screen, or the exact UI element under your cursor. Each one lands in your narration as a numbered reference, right where you said it.
- ✎ Add selection — captures the focused app’s selected text
- ▣ Grab region — freeze the screen, drag a box, it’s in the note
- ⌖ Pick element — outline what you point at, click to capture it with its text
Finish, and Deixis writes the narrative: your words with the evidence inline. A code review, a bug report, a design walkthrough — spoken in one pass instead of assembled from screenshots afterwards. All of it stays on your machine.
“The settings panel crowds the mic picker — this label wraps at normal widths…”
⌖ ref 1 · element · “Input device” combo, 312×46
“…and the save row disappears below the fold. Here’s the whole panel at 125% scaling:”
▣ ref 2 · grab · region 840×620
“The fix is the same clamp we used in the history view — this is the code I mean:”
✎ ref 3 · selection · max-height: clamp(18rem, 60vh, …)
One chord, every mode
Hold Ctrl+Win and speak — or tap a letter
Verifiable, not claimed
Unplug your network cable. Deixis keeps working.
Recognition is in-process
Whisper runs inside the app on your CPU. There is no speech API, no key to configure, and nothing to send your voice to.
There is no account
No sign-in screen exists in the app. Install it, hold the key, dictate. Identity only ever enters the picture for optional sharing features — and those are off by default.
No telemetry
The default configuration makes no network calls at all. The optional cloud cleanup is one endpoint, off until you turn it on — and you can point it at your own key instead.
Local doesn’t mean slow
Measured on the reference machine (desktop CPU, RØDE PodMic USB), release build,
small.en model. Methodology ships with the app: run deixis-core --bench yourself.
| Path | Latency | What it means |
|---|---|---|
| Short utterance, speech→text | ~100 ms | a command or a sentence fragment |
| Full dictation, key-up→text inserted | < 500 ms | you release the key and the text is there |
| Network round-trip we don’t make | 0 ms | a cloud hop alone costs 500 ms+ before recognition starts |
One tool, not four subscriptions
Everything you say becomes one private corpus — on your machine
Screenshots & annotation
Every monitor frozen the instant you press Ctrl+Win+S — so menus, tooltips and hover states survive. Select a region, a window, or an element; annotate with arrows, boxes, text, highlights and numbered steps; copy or save. The screenshot tool you own, not rent.
Meetings v1
Record in-person or online meetings — your mic plus system audio, no bot joining the call. Timestamped transcripts with speaker labels and AI meeting notes, all generated locally. Your meetings never leave your machine.
Screen video v1
Record your screen and voice, get a link to share — the viewer needs nothing installed. Recording is local and free; hosting and share links are what Pro is for.
A dictionary you own
Tuned for developer vocabulary out of the box — useEffect,
TanStack, shadcn come out right — and a per-user dictionary
teaches it your project’s names, with find-and-replace corrections for the ones the model
refuses to learn.
Context-aware cleanup
Deixis reads which field has focus and formats for it: prose in a doc, a commit message in your terminal, a message in your chat. Rules first — instant and local — with optional LLM cleanup you can run through Ollama on your own machine.
Hands-free notes
Double-tap the chord and it records until you stop — a meeting-length voice note becomes a reviewable transcript, never text dumped into whatever window had focus.
Loom, Wispr and Granola are three separate clouds; they structurally cannot assemble what this builds: one local, searchable corpus of everything you said — in meetings, in dictation, over your screen — under one dictionary and one hotkey.
Pricing
Local is free. Forever. That’s the point.
Free
$0
Everything local. No account. Forever.
- Dictation, dictionary, cleanup
- Free mode — talk, point, capture
- Screenshots & annotation, translate
- Hands-free notes & history
- Meetings: transcripts, speakers & notes, locally v1
- Local LLM cleanup & notes via Ollama
Pro coming soon
$49/year
or $10 a month, if you prefer.
- Screen video with share links
- 100 GB hosted storage
- Managed cleanup & meeting notes — for machines that can’t run a local model
Team coming soon
$39/seat/year
Less per seat than Pro — deliberately.
- Everything in Pro
- Shared dictionary & admin policy
- Pooled storage
Replaces $363/yr of tools — Loom Business ($180) + Wispr Flow ($144) + Snagit ($39) — at 86% less. And no feature that runs on your machine will ever move behind the subscription.
Try it in the next five minutes
Install, hold Ctrl+Win, speak. No terminal, no config, no account.