01 Operating lane
Dictation overlay — listening · live waveform · local whisper
macOS · VOICE CONTROL LAYER
Speak in. Code out.
Dictate, format, ask, remember, and dispatch work — from one always-available macOS overlay. Not voice typing. A voice interface to your tools and your agents.
structured intent ● local whisper
Analyze the repo for duplicate event-dispatch — propose a minimal patch.
handed to your agent · run started
NOT A WINDOW YOU SWITCH TO
It lives over your work.
Summon the overlay with a hotkey on top of whatever app has focus — your editor, the browser, TextEdit. Speak, and the result lands right where you're working.
ONE INTERFACE
Every way you want to speak.
DICTATION
“just get this thought down before I lose it”
pasted into your editor · live preview
Dictation
Hold Fn, speak. Live preview, then it pastes into the app you're in.
Formatting
An AI pass cleans punctuation and structure — your meaning untouched.
Agent Chat
A persistent conversation — ask, refine, attach context, come back later.
Selection-aware
Highlight anything, say what to do, get the result back in place.
TWELVE THAT MATTER
Not voice typing. A voice control layer.
Twelve surfaces that separate codescribe from a generic STT box — each one is in the product today, not a roadmap poster.
-
Multi-selection context buckets
Combo hotkeys during dictation capture multiple selections into context for agentic chat and daily transcript flows — not one clip, not one paste.
-
Apple live freezed + append
Multi-utterance live assembly: previous finals freeze, current interim appends. Long Polish thoughts stay coherent instead of collapsing to a tail.
-
Live × Whisper merge fill
Apple is the floor of truth; Whisper fills gaps. Full-replace is forbidden — correct tokens are never deleted by a second pass.
-
On-device Polish as foundation
SFSpeechRecognizer on-device for pl-PL is a product path, not a legacy fallback. Speech never has to leave the Mac for live dictation.
-
Overlay over any focused app
Global hotkeys summon a dictation surface on top of whatever owns focus — editor, mail, browser, terminal — without a mode switch.
-
Hold · toggle · hybrid triggers
Pick how recording starts: press-and-hold, toggle, or hybrid. Hands-free silence windows when you need them; snappy hold when you don’t.
-
Selection-aware rewrite in place
Highlight text anywhere, speak the intent (“make it formal”), and the result lands back in the same field — no copy-paste theatre.
-
Agentic chat + MCP tools
Basic dictation or full agentic lane: persistent chat, tool substrate, and voice that can dispatch work instead of only typing it.
-
Teacher dictionary that learns
Corrections become live lexicon rules. Teach mines the store into variant→canonical pairs the post-processor actually applies.
-
Honest runtime truth
Permission matrix, Active STT status, and Speech Recognition TCC are live probes — not static “looks fine” badges that lie.
-
Local Whisper when you want it
Slim daily build ships Apple live without a gigabyte model. Whisper downloads from Settings when you want file-final fill or offline paths.
-
First-run that actually unlocks speech
Setup wizard includes Speech Recognition with the same grant path as Mic and Accessibility — so Apple live works after install, not after a support ticket.
IT'S NOT JUST TRANSCRIPTION
Messy out loud. Clean on the page.
RAW SPEECH
dobra zróbmy tak najpierw sprawdź ten plik potem zobacz testy i nie rozwal scope
FORMATTED
Dobra, zróbmy tak: najpierw sprawdź ten plik, potem zobacz testy — i nie rozwal scope.
Punctuation, capitalization, paragraphs, artifacts removed. Meaning kept exactly.
WORKS OVER ANY APP
Select. Speak. Done.
Highlight text anywhere, say what to do, and the result drops back in its place — no copy, no chat window, no context-switch.
Mail — reply draft
Hey, yeah we can do the call. i think tuesday works but maybe later is better idk, lmk what suits u Talk soon.
Mail — reply draft
Hey, yeah we can do the call. Tuesday works for me, though a later slot would be preferable. Please let me know what suits you. Talk soon.
"fix the grammar" "summarize this" "translate to Polish" "explain this code" "turn into a checklist"
CONTINUITY
A persistent agent,
not a disappearing
prompt box.
Talk to an agent that stays with your work. Ask, refine, attach files, follow up — and restore a thread when you need it again.
WORKING HISTORY
Your voice doesn't disappear.
PERSONALIZATION
Make the agent yours.
Edit the prompts that drive formatting and assistive mode. Your voice. Your rules. Your agent.
formatting.prompt edit · save · reset
Keep my meaning exactly. Short sentences.
Prefer bullet points for lists. Never touch
code blocks. Always keep the original language.
assistive.prompt edit · save · reset
Return only the result, no preamble. Respect
scope — don't expand the task. Be operational.
Match my tone. Output a ready-to-run patch.
AGENT-STACK READY
Speak tasks into your agent stack.
Native tools and MCP support turn spoken intent into structured input for agent runtimes — a voice control layer for VibeCrafted-style workflows.
clipboard selection screenshot file read type text GitHub fetch thread search + MCP tools
power-user · dispatch "run a repo check and prepare a minimal patch plan."
Your voice becomes
the control layer.
Not a tool you open. A layer that sits over everything you already do — and turns talking into doing.
Download for macOSBUILT FOR macOS · APPLE SILICON
A real native app. No web wrapper.
A menu-bar tray, a floating overlay, and a settings window that tells you the truth — local-first STT, permissions, and runtime state at a glance.
Menu-bar tray — show agent, history, notes, diagnostics
Shortcuts — one gesture per mode, applies live, no restart
Keys — stored in the macOS Keychain, write-only, never shown back
SETUP · 12-STEP WIZARD, SIX MOVES
Onboarding in six moves.
First launch to a ready voice runtime — pick a lane, grant permissions, set language, connect a provider, choose hotkeys, check readiness.
02 Screen recording
03 Dictation language
04 AI provider
05 Trigger hotkeys
06 Agentic readiness
— from first launch to a ready voice runtime —
macOS · APPLE SILICON · v0.13.0
Local-first. Cloud opt-in.
Download Codescribe.dmg signed & notarized · 1.4 GB · build from source: see README