Captures both sides
Your mic and your computer's audio — Zoom, Meet, a Slack huddle — recorded as two clean, separate streams. "You said" vs "they said" stays unambiguous, and no bot has to join the call.
Humla records your mic and the other side of the call — no bot joins the meeting — transcribes it, labels who said what, and writes a clean summary. Run it on OpenAI, Deepgram, Groq, or fully on your Mac.
No sign-up to start. Bring your own API key, or record completely offline.

Transcribe with your choice of engine
Made for people who'd rather their conversations didn't live on someone else's server.
Your mic and your computer's audio — Zoom, Meet, a Slack huddle — recorded as two clean, separate streams. "You said" vs "they said" stays unambiguous, and no bot has to join the call.
No backend, no telemetry, nothing phones home. Run local Whisper plus a local model and no audio or text ever leaves your machine. Keys live in the macOS Keychain, notes in a local database.
Humla feeds the model your typed notes for intent and the transcript for facts, then writes the shape you pick — Meeting, 1:1, interview, and more.
Three steps, and every one of them is a setting you control.
Humla captures your microphone and the meeting audio together, in one click. Pause and pick back up whenever — nothing is sent anywhere until you stop.
Pick an engine per language — say Norwegian on local NB Whisper, English on Deepgram. When you stop, an on-device pass labels each speaker; click a name to rename it everywhere.
Choose a preset and Humla writes clean Markdown in your language — ready to tidy up in a keyboard-driven editor, or drop straight into your docs.
Two free local engines label every turn and adapt to how many voices are in the room — no audio uploaded.
Set a default, then override by language. Humla routes each recording to the right provider automatically.
Bias names, jargon, and acronyms into the transcription so they spell consistently every time.
Search across titles, note bodies, transcripts, and folders — find that decision from three weeks ago in seconds.
Coloured speaker pills inline; click anywhere to fix a word. Playback highlights the transcript as it plays.
Gatekeeper accepts it directly, and installs quietly update themselves on launch.
Humla is local-first and works fully offline. When you want to sync across your devices or share workspaces with teammates, point the app at a sync server. Transcription and summaries still run through your own providers or on-device — the server only syncs, it never sees your keys.
Start a note on your laptop, finish it on your desktop. Notes, folders, transcripts, audio, and summaries stay in step.
Give the whole team one place for its meetings, with shared folders instead of notes scattered across laptops.
If you sit in agency meetings under NDA, Humla is built the way you'd want it to be.
Clients never watch a recorder join the call. You're simply recording your own machine — nothing appears on their side.
Local Whisper for transcription plus a local model for summaries means no audio or text ever leaves the Mac. Ideal for sensitive calls.
Provider API keys are stored by the operating system, one entry per provider — never in plaintext on disk.
Humla doesn't phone home. The only outbound traffic goes to the provider endpoints you explicitly configure.
Run team sync on your own box or in-region storage, so client recordings stay wherever your compliance needs them.
The app is MIT licensed and public. You — or a security team — can read exactly what it does before you trust it with a call.
And even then, only if you'd rather not run the server yourself.
The full desktop app on a single Mac. No account needed.
Hosted team sync for everyone in the workspace. Nothing to run.
The same team sync, running on a box you control.
No. Humla records the meeting audio locally through macOS itself, so there's no participant to admit and nothing for the other side to notice. You're just recording your own machine.
The defaults keep everything on your machine. Humla has no backend and no telemetry. If you point it at OpenAI, Deepgram, or Groq, only the audio you record goes to that provider under your own account. Use local Whisper plus a local model for summaries and nothing leaves your Mac at all.
Only if you want a cloud transcription engine. On-device Whisper and speaker ID run with no key and no cost — a one-time model download and you're set.
It syncs your notes and shares workspaces across a team — that's all. It does not transcribe or summarise for you; that always runs through your own providers or on-device. It's a hosted sync server so you don't have to run one yourself.
Yes. Humla's sync engine talks to a PocketBase backend, so you can point the app at your own PocketBase server — hosted or running locally — instead of Humla Cloud. Self-hosted servers have no subscription gate: every team feature is free. The subscription only exists on the managed Cloud option.
Effectively any language Whisper supports — around 100, with automatic detection — plus a Norwegian-tuned NB Whisper model that runs locally. You can even route different languages to different engines automatically, so Norwegian goes to local NB Whisper while English goes to Deepgram.