Advice lands while the call is still live
A client drops a number that sounds off. Before you finish typing a note, the fact-check card is in the side panel — with a suggested follow-up. By the end, the action items are already written.
VoxAI · The missing sensory layer for AI
Give your AI ears —
and a voice.
It sits in your meetings and knows who said what. It slips you advice while the call is still live. And when you talk to it, it talks back — out loud. All of it can run entirely on your Mac.
Free public beta · Notarized · macOS 14+ · Apple Silicon
Connected — live dialog
01 — A day with VoxAI
Four moments it pays for itself.
Every line lands under the right name
Speakerphone with a vendor: mic is you, system audio is them — labeled in real time, no voiceprint guessing. Hang up, and the named transcript is just… there.
Your build reports back — out loud
Tell Claude Code what to ship, from across the room. When the tests pass, it says so. You never walked back to the screen.
"What did we decide?" gets a real answer
Ask about pricing from last Tuesday. Your AI reads the speaker-labeled transcript, answers, then drafts the recap and to-dos — ready to paste into the team channel.
02 — How it works
Ears. Advice. Voice.
Your AI is smart but senseless: it can't hear the meeting you're in, and it can't speak up when it matters. VoxAI is the missing layer — a native macOS recorder-transcriber with an embedded MCP server.
01
Ears
Live speaker-labeled transcripts stream to your AI while the meeting is still happening.
02
A voice in the room
It pushes advice cards into the meeting window, silently: fact checks, risks, negotiation strategy.
03
A voice out loud
speak(text) plays through your Mac; the floating dialog turns your speech into its input — a full two-way voice conversation.
03 — What's in the box
Everything a copilot needs.
The meeting copilot
Every line gets a speaker's name — live, while you talk. Your AI checks in by name, follows the growing transcript, and drops color-coded notes in the side panel: follow-ups, action items, risks, strategy. Export Markdown with the notes attached, or hand the whole thing over for a summary.
Call mode: channel is identity
System-audio capture puts the far side on its own channel — mic is you, system audio is them. Real-time labels with zero guessing; no voiceprint model in the loop.
A voice dialog with your AI
A compact chat window: your speech goes straight to the AI; replies come back as bubbles — and out loud, if you want. Light or dark, following your system.
Refinement that outlives the app
After the meeting, a second pass sharpens everything: mlx-whisper re-transcribes on the Apple GPU, pyannote splits speakers word by word. It runs detached — quit the app, the job still finishes. One-click undo if you preferred the original.
Connect your AI in one click
Settings detects what's installed — Claude Code, Claude Desktop, Cursor, Codex — and generates the exact config. The MCP server ships inside the app: no Python, no venv, nothing extra to install.
04 — Local-first
Nothing has to leave your Mac.
Recording, transcription, speaker labels, and the flagship voice all run on-device. In a blind listening test, the local Qwen3 voice held even with the cloud — six voices, no account, nothing uploaded.
- Local voice — Qwen3 1.7B, near-cloud quality, six voices; a one-time ~2.2 GB download, with the system voice filling in until it lands.
- Local ears — live transcription and speaker labels on the Apple GPU; post-meeting refinement by mlx-whisper large-v3-turbo.
- Cloud, strictly opt-in — a free cloud voice, Azure with your own key, cloud refinement: each one is off until you switch it on, behind an explicit upload notice.
No ports, no telemetry, no account — the MCP server is local stdio.
05 — Works with your AI
Two tiers, twelve tools.
Any MCP client gets the archive; an agentic one becomes a copilot. The server writes its full protocol into MCP instructions — a brand-new AI learns the rules the moment it connects.
Archive
- needs
- any MCP client — Claude Desktop, …
- you get
- browse & summarize recordings, speaker-labeled transcripts, TTS
Copilot
- needs
- an agentic client that can loop — Claude Code, …
- you get
- live listening, real-time advice, voice-direct dialog
list_meetings() All recordings with metadata and live status get_meeting(id) Full speaker-labeled transcript get_live_transcript(after_index) Incremental live transcript + speaker map push_advice(text, kind, label?) Advice card into the meeting window start_listening(name) / stop_listening() Copilot check-in / check-out get_dialog_input(after_index) Your speech as chat input post_dialog_reply(text) AI reply bubble in the dialog window speak(text, voice?) / stop_speaking() TTS out loud / interrupt list_voices() / update_voice_config(…) Voice management 06 — Get it
Put a voice on your desk.
One notarized .dmg. Drag it to Applications, connect your AI in a click — and your meetings stop being write-only.
Download .dmg — coming soon
Free public beta v0.9 · Notarized (Developer ID) · macOS 14+ · Apple Silicon