Everything it does while someone is watching you
Each piece exists to solve a problem that showed up in a real call - headphones breaking the microphone, answers arriving too late, a window appearing in a screen share.
Token bucket is the usual answer. Each client gets a bucket that refills at a fixed rate:
- Refill r tokens per second, capped at burst size b.
- A request costs one token; an empty bucket returns 429.
- Store the count and last-refill timestamp in Redis.
It absorbs short bursts, which a fixed window cannot
Twelve things, no filler
System audio capture
Hears what plays on your computer through ScreenCaptureKit, so headphones change nothing and your microphone stays free.
Live transcription
Deepgram Nova-3 turns the other person's speech into text as it is spoken, with an interim line you can watch fill in.
Complete questions
Waits for the finalised transcript before answering, so a trailing 'and give one example' is not lost.
Streaming answers
Words appear as DeepSeek writes them. Nothing sits behind a spinner while a full answer is assembled.
Thorough answers
Tuned for full explanations with structure and examples, not the two clipped sentences a default chat prompt gives.
Screenshot questions
Capture the screen and ask about it. A diagram, a stack trace or a shared document is answered directly.
Markdown rendering
Headings, lists and fenced code blocks, sanitised before they reach the page so nothing executable slips in.
Hidden from capture
Hide mode excludes the window from screen sharing and recordings while it stays on your own display.
Always on top
Floats above the call with adjustable transparency, so you can read through it to whatever is behind.
One-key recording
Spacebar or the microphone button starts and stops listening. Everything is on a single row of controls.
Fast first word
Reasoning is switched off by default and the connection is warmed before the question lands.
No keys to manage
Sign in and go: answers and listening run on your plan's credits. Your resume and knowledge base stay on your computer.
It hears the meeting, not you
The audio playing through your speakers or headphones is what gets transcribed, so your microphone stays free for the conversation.
- Captured through ScreenCaptureKit - headphones change nothing
- Deepgram Nova-3 transcribes as they speak
- Waits for the finalised text so trailing clauses are not clipped
So how would you keep the cache consistent across regions?
And what breaks first under load?
Ask about whatever is on the screen
One click sends the current screen with your question, so a diagram, a stack trace or a shared document can be answered directly.
- DeepSeek reads the image and the question together
- Resized before sending, so it stays fast
- Answers come back as Markdown - code stays code
Invisible to whatever you are sharing
Hide mode keeps the window off the shared stream while it stays on your own display, floating above the call.
- Excluded from screen capture, not just moved aside
- Always on top, with adjustable transparency
- Keys live in a local file - nothing is proxied through us
Hide mode keeps the window out of the shared stream while it stays on your own display.