Answers in the room
One line, phrased to be said out loud, with the reason underneath. Latency to first word is measured and shown, so you know what you are trusting.
Umbra sits beside your call, hears both sides, and puts the sentence you need on screen in about a second. When the call ends, the recap and the follow-up email are already written.
No bot joins the call. No one else hears about it. Umbra listens to the audio already playing on your machine, answers while the question is still open, and closes the loop before you stand up.
⌘⇧L opens a channel to the system audio and your microphone, separately, so “them” and “you” stay distinct. A listening badge sits in the pill while it runs.
⌘⇧A answers what was just asked. Or turn auto-answer on and the sentence is already there when the question lands — one line you can say out loud.
Decisions, owners and dates are captured as they are said. The recap and a ready-to-send email are waiting in the dashboard when the call ends.
One line, phrased to be said out loud, with the reason underneath. Latency to first word is measured and shown, so you know what you are trusting.
Every call becomes a meeting file: transcript, answers, notes. Search across all of them from the dashboard instead of scrolling a chat thread.
A recap plus a sendable email, drafted from what was actually said. Copy it, edit the two words that are wrong, send it before the next meeting starts.
⌘⇧R freezes the screen and lets you drag a box around a deck, a ticket or an error. Only that box is sent, cropped at native resolution.
OpenAI, Anthropic, Google, Groq, or Ollama on your own machine — switch provider and model in two clicks. A test button reports the real error instead of a shrug.
No account, no telemetry, no sync. Keys, transcripts and notes sit in one folder you can open and delete. Nothing leaves the machine except the request to the provider you chose.
Offline mode refuses every request before a socket opens. Pair it with Ollama and a local model and the whole loop — transcription aside — stays on the machine.
The panel sets content protection, so when you share your screen the usual conferencing apps capture a blank area instead of your notes. Best effort — measured by the app, not promised by marketing.
On first launch, with nothing configured, Umbra still answers by pulling the strongest lines out of what was said. It says that is what it did, every time.
There is no Umbra account, because there is no Umbra server holding your data. Settings, meetings and model keys all live in one folder on the machine in front of you.
If you use a hosted model, the audio transcript and the region you selected go to that provider — that is the deal you are making with your own key. If you use the included model, they go through our proxy instead, counted by the hour and not stored. Offline mode refuses both.
Read the honest limits →No installer wizard to click through, no account to create, no key to paste before your first answer. Add a hosted model later if you want longer answers.
Node 20 or later, then three commands. The same three produce a macOS or Windows installer.
npm install
npm run smoke # boots it, runs one ask, exits non-zero on failure
npm run dist:mac # or dist:win
On Windows, Ctrl replaces ⌘.
Bring a model key, or bring nothing and use the local engine — either way there is no seat count, no trial clock, and no activation server to be offline from. Paying is for the convenience of not managing a key.
The included model is metered by meeting hours, because transcription is what costs money — not by answer. Checkout opens in your browser and the licence key is pasted into the app; the app itself never needs a login.
Every tool has edges. These are the ones that would annoy you if you found them later.
No. Nothing is admitted to the call and no participant list changes. Umbra reads the system audio already playing on your machine, plus your microphone on a separate channel, so it knows who said what.
No. With no key, Umbra runs its local engine: instant, extractive, honest about being extractive. Add an OpenAI, Anthropic, Google or Groq key — or point it at Ollama — when you want a model that reasons.
One JSON file per meeting in ~/Library/Application Support/Umbra/data. Settings sit beside
them. Delete the folder and everything is gone — that is the whole backup story.
Only if you tell them. The panel keeps itself out of your screen share and stays out of the participant list, so nothing announces it. That makes disclosure your decision, and we think you should make it.
If the interviewer knows, yes — it is a good cue card. There is no mode for pretending you are not using it, and we are not adding one. For the interviewer's side, the Recruiting mode is built for probing depth and taking notes.
There is nothing to log into. Downloads are plain files, settings live in the app, and a licence key is pasted into the app rather than tied to an account here.