Record
Hold the Fn key and Lavox records your voice in any app. Connect Google Calendar and each recording is named after the meeting it belongs to.
Meeting recorder · dictation · memory
Every meeting, understood.
Lavox listens, transcribes and remembers on your Mac: every speaker, every decision, even the option you turned down. Any AI you use can recall it. Nothing gets buried.
Open source · MIT · on GitHub The bar
Always on. Never in the way.
The bar is a thin glass pill under your notch. Hold the Fn key and it records, dictates or takes notes in whatever app you're in.
fn hold to talk, anywhere
Dictation
Meeting soon
Video soon
AI Note soon
Transcription
Who said what, and what was decided.
Record from the bar and Lavox transcribes on your Mac, naming every speaker. Who said what, what was decided and what is still open, all of it searchable.
Weekly Product Sync · recorded from the bar
The memory
Ask in August what you decided in March.
Meetings and dictation land in one memory on your Mac: decisions, the options you turned down, commitments, facts. Any AI you use can read it over MCP and answer with the exact quote.
The memory loop · live on your Mac
01 · You speak
02 · Lavox remembers
Fact
The beta ships behind a waitlist
Commitment
Demo video due Friday
Decision
Postgres for the sync store
anchored to the verbatim transcript 03 · Any AI recalls answered from
Decision from Aug 19: Postgres over Redis, because Redis alone wouldn't survive a restart.
“Redis alone won't survive a restart.” verbatim · from your recording
lavox-memory · via MCP Works today with Claude Code · Claude Desktop · any MCP client
How it's stored
verbatim layer
The transcript is canonical. Speaker-turn chunks with context headers:
[date | kind | title | speaker] + text assertion layer
A typed index extracted on top, never a replacement. Three columns make it bitemporal:
occurred_at · invalidated_at · superseded_by
One SQLite file in ~/Lavox/memory. Embeddings run locally (ONNX), and
the model is swappable.
Why it's built this way
The full design, with the measurements behind it →
Extraction is an index, not the record.
Verbatim chunks beat lossy fact-extraction on recall, so both layers are searched and fused. Every assertion links back to the exact chunk it came from.
Corrections supersede, they never delete.
Change your mind in August and "what did we believe in July?" stays answerable. History is a schema, not a graph database.
Nothing enters anonymously.
Every write carries a source: extracted with a transcript anchor, stated by you, or written by an agent. That is the defense against memory poisoning.
The whole thing is open source.
Recorder, transcription, memory and MCP server, all under the MIT license. Read the code, run it on your own hardware, swap in your own models.
How it works
Three levels down to the truth of a meeting.
No bot joins your call and nobody gets stuck taking notes. Lavox records where you are and descends through the conversation so you don't have to.
Transcribe
Whisper transcribes on your Mac by default, and speaker recognition separates and names every voice.
Recall
Export clean markdown into Obsidian, or just ask your AI: the built-in memory answers over MCP, quote included.
Benefits
Recorded where you are. Kept on your Mac.
Local-first, private
Transcription runs on your Mac by default, so your voice never has to leave it. Anything that touches the cloud runs on keys you bring yourself.
Every voice named
Speaker recognition separates the voices and tells you who said what. Enroll a voice once and it is labeled by name from then on.
Your notes, your vault
Summaries and transcripts export as clean markdown straight into Obsidian, and your AI reads the same memory over MCP.
0 bots joining your calls recording happens on your Mac
0 capture modes on the bar dictate · meet · video · note
0 key to record anywhere hold fn · push to talk
0% local transcription by default your audio stays on your Mac
FAQ
Questions from the surface.
Which meeting apps does Lavox work with?
All of them. Lavox records from your machine through the bar. It doesn't join calls as a bot, so Zoom, Meet, Teams or a hallway conversation all work the same way.
Is my audio private?
Transcription runs on your Mac by default, so recordings never have to leave it. Optional cloud features run on keys you provide, and your data is never used to train anyone's models.
How does speaker recognition work?
Enroll a voice with a 20-second sample. From then on, Lavox separates the speakers in any recording and labels them by name.
What platforms does the desktop app support?
macOS 13 or later on Apple Silicon, with the bar living under your notch. Windows is on the roadmap.
What is the memory, and which AI tools can use it?
Everything you record or dictate is distilled into a structured memory on your Mac: decisions with the rejected alternatives, commitments, facts, each anchored to the verbatim transcript. A built-in MCP server exposes it to Claude Code, Claude Desktop and any other MCP client, so your AI answers from what you actually said.
Is Lavox open source?
Yes. The recorder, the local transcription, the memory and its MCP server are on GitHub under the MIT license. You can read every line, self-host the sync server and bring your own models.
Does it integrate with Obsidian?
Yes. Every recording exports as clean markdown with frontmatter: summary, tasks and transcript, straight into your vault.
Get started
Let the meeting take its own notes.
Lavox records where you are, transcribes on your Mac, and hands you the summary minutes after the call.
macOS app · open source · MIT