macOS and Windows · speech transcribed on your own machine

The answer, while the question
is still being asked

A desktop app that answers technical interview questions as they are asked. It hears the question, answers in about two seconds, and does not appear in screen sharing.

Get started free Download

30 minutes of credit when you sign up. No card.

✕– ‹› 1/1 What happens when two threads write to the same HashMap? ◴ 0:00 ✕
Hear⌃⌥L Mic⌃⌥I Answer⌃⌥A Screen⌃⌥S

Press Screenshot (⌃⌥S) to answer what is on screen.

AA ⋯
  1. It listens, and shows what it heard The strip along the top is the live transcript, made on your own machine.
  2. You press one key ⌃⌥A. It works out which part of the last minute was the question.
  3. The answer starts in about a second It streams — you read while it is still being written.
  4. Short answer first, then the detail Shaped to be said out loud — the one-line answer, then what backs it up.

The panel above is the real interface, rebuilt here in the page so the text stays sharp at any size. Screenshots and your own voice work the same way — one key each. All three inputs →

What it is

A small window that floats over whatever you are doing, and answers technical questions while an interview is happening.

How it hears

Three ways: a screenshot of the question, the audio of the call, or you asking out loud. Speech is transcribed on your own machine.

What it costs

Credit by the hour, spent only while a session is running — or an unlimited pass. Thirty free minutes to try it, no card.

And before the day

The whole NeetCode 150 worked in Java, a DSA cheatsheet and an AWS course — free, no account needed.

Inputs

Three ways it hears the question

One key each. All three end up in the same place.

Answer⌃⌥A
Priya asks — press once and it answers what it heard.

From the conversation

The call's audio, transcribed on your own machine as it goes. Never uploaded.

Screen⌃⌥S
A question on screen — one press captures and answers it.

From your screen

Whatever question is on screen — a coding pad, a shared doc, a take-home.

Mic⌃⌥I
Say it out loud and it answers you back.

From your own voice

Ask out loud when you want to think something through.

Not the answer you wanted? ↻ re-asks for a different angle, ⌃⌥F follows up on the answer in front of you — so “give me the optimised version” works without repeating yourself.

New

The answer, as a picture

You cannot read a paragraph while an interviewer waits. When an answer has a shape — a request path, an exchange between services, a sequence of steps — the app draws it and puts the detail underneath.

Any question with a shape gets one

Two forms, chosen by what the answer actually is: a flow for a path something travels, and a sequence for an exchange between parts of a system. Comparisons stay as tables and code stays as code — both are already read faster that way, so neither is forced into a diagram that would only slow it down.

  • Drawn, not pastedNative, so it stays sharp at any window size and any opacity — not an image dropped into the panel.
  • Labels you can sayEach arrow carries a few words, so you can talk through the picture without reading from it.
  • Only when it helpsA diagram that repeats the bullets is noise. The model is told not to draw one unless there is a real shape.
System designWhat happens when a user clicks a shortened URL?
Shortened URL request
BrowserCacheOrigin
GET /aB3x
302, full URL
GET the page

A hit never reaches the origin, so the redirect costs one round trip.

TechnicalWalk me through the write path of a message queue.
Write path
ClientLoad balancer API serverMessage queue WorkerObject store

The write returns once the queue has it; the worker drains it afterwards.

Privacy

Nobody on the call can tell

Six of these the app proves to itself on every build. The seventh is the one everybody overstates, so it says what is actually true. Open any of them.

Screen share

Share your screen. The panel is simply not in what they receive.

verified

Invisible on screen share

Share your screen. The panel is simply not in what they receive.

The window asks the system to exclude itself from capture — on macOS through both of its screen-capture paths, on Windows through display affinity — so it is never drawn into the stream, not layered behind something, not sent at all. Every sharing app reads the screen through one of two system paths, and the app captures itself through both on every build.

checked every build
✕– ‹› 1/1 What happens when two threads write to the same HashMap? ◴ 0:00 ✕
Hear⌃⌥L Mic⌃⌥I Answer⌃⌥A Screen⌃⌥S

Press Screenshot (⌃⌥S) to answer what is on screen.

AA ⋯
What they seeWhat you see

The Dock

No icon, no bouncing, no entry while it runs.

verified

Invisible in the Dock

No icon, no bouncing, no entry while it runs.

A background agent: LSUIElement in the bundle and .accessory at launch. There is no tile to notice and nothing to right-click.

checked every build

Activity Monitor

Search the process list for the product and nothing comes back.

the honest one

Nothing called BunnyBuddy in Activity Monitor

Search the process list for the product and nothing comes back.

The process has a different, unremarkable name. Being straight about the limit: no app can truly hide from Activity Monitor, and one that claimed to would be doing something you should not want installed.

what is actually true
Activity MonitorBunnyBuddy
No matching processes

Tab & focus checks

Using it does not blur the page you are being watched on.

verified

Invisible to tab and focus checks

Using it does not blur the page you are being watched on.

Proctoring software watches for the moment its window loses focus — a blur or visibilitychange is all it takes. The panel is non-activating: clicking, dragging and resizing it never take focus, and it is absent from ⌘-Tab.

checked every build
✕– ‹› 1/1 What happens when two threads write to the same HashMap? ◴ 0:00 ✕
Hear⌃⌥L Mic⌃⌥I Answer⌃⌥A Screen⌃⌥S

Press Screenshot (⌃⌥S) to answer what is on screen.

AA ⋯

The cursor

No crosshair, no resize arrow, no hint that anything is there.

verified

Your cursor never changes

No crosshair, no resize arrow, no hint that anything is there.

The pointer is drawn by the window server and is captured even though the panel is not — so a cursor that changes shape over empty desktop is the one thing the exclusion cannot hide. The build fails if any code touches it.

checked every build
✕– ‹› 1/1 What happens when two threads write to the same HashMap? ◴ 0:00 ✕
Hear⌃⌥L Mic⌃⌥I Answer⌃⌥A Screen⌃⌥S

Press Screenshot (⌃⌥S) to answer what is on screen.

AA ⋯

Your audio

Speech becomes text on your own machine, about 10 ms behind.

verified

The audio never leaves your machine

Speech becomes text on your own machine, about 10 ms behind.

A speech model downloads once and then runs locally. The call’s audio is never uploaded to anyone — only the question text is sent, and only when you press a key.

checked every build
the call→your machine10 ms behind⇢anywhere elsenever

No extension

Nothing is installed into the meeting or the assessment page.

verified

There is no browser extension

Nothing is installed into the meeting or the assessment page.

Most tools here ask to be added to your browser, which puts their code inside the page being proctored, where the page can enumerate it. This is a separate window looking at your own screen.

checked every build
Extensions
Password Manageron
Grammar Checkon
—nothing from us
The obvious question

Why not just switch to a chat window?

You can. Here is what happens when you do.

Another windowBunnyBuddy
In the screen share Visible — it is a normal window Excluded at the window level
Getting the question in Type or paste it yourself Already heard it; one key
Where you are looking Away from the call, into another app A panel over what you were doing
What it knows about you Whatever you re-explain each time Company, role, JD and résumé, once
Shape of the answer An essay to skim mid-sentence One line to say, then the detail
Speech recognition

so tell me, what happens when two threads write to the same HashMap

Transcribed on your own Mac, 10 ms behind

The speech model downloads once and then runs on your machine — no round trip to anyone's server, which is why it keeps up. The call's audio is never uploaded; only the question text, and only once you press a key.

Answers
Speech → text, as they talk10 ms
You press ⌃⌥A0 ms
Finished, streaming all the way3.4 s

10 ms behind the speaker, start to finish in 3.4 s

The transcript runs about 10 ms behind the audio — measured on a Mac, median over a spoken question, and the app fails its own build if that passes 120 ms. So by the time you press a key the question is already text: none of the wait is speech recognition. What is left is the model composing the answer, and it streams, so you are reading it while it is still being written.

Compatibility

It works with whatever they are using

Staying out of the screen share is not an integration with any one app. The window asks macOS to exclude itself from capture, and every screen-sharing app takes its pixels through a system path the window excludes itself from. The app tests both on every build and refuses to ship if it shows up in either.

The one thing that beats it is a second camera pointed at your screen — nothing running on the Mac can help with that.

Why it works mid-call

Built for the two seconds you have

⚡

About two seconds

First words in roughly two seconds, streaming as they're written.

🔒

Speech stays on your machine

Transcription runs locally. Only the question text ever leaves, once you ask.

◱

Discreet by design

Never takes focus, excluded from screen capture, one key to hide it.

✎

Grounded in your background

Company, role, job description and résumé once — answers cite your actual experience.

☰

Answers shaped your way

Brute force then optimised, comments on every non-obvious line — your rules win.

⤓

Every session kept

Every question and answer saved, searchable, exportable as Markdown.

Learning

The weeks before the interview

Everything above helps during the interview. Learning is for the weeks before it — built from what companies are actually asking, not from a textbook written five years ago.

Live now · free

DSA, worked the way an interviewer wants to hear it

The NeetCode 150 in Java — but not a solutions dump. Every question is walked from the constraints to the pattern, to brute force, to the optimal solution, then handed back as a script for saying it out loud. Plus a complete cheatsheet of every structure and Java mechanic the 150 require.

Start reading →
150questions
23cheatsheet topics
Freeno account needed
Live now · free

AWS for backend developers, zero to senior

A 14-day course taking you from no AWS knowledge to designing, building and defending production AWS-backed systems at a senior-engineer interview — identity, networking, containers, serverless, relational data and DynamoDB. Not a certification cram.

Open the course →
7chapters
7written in full
Coming soon

Real interview questions

Questions actually asked at real companies, grouped by company and topic, with how often each comes up. Contributed by the people who were asked them.

FAQ

Questions

What do you do with my audio?

Nothing leaves your machine. Speech is transcribed locally by a model bundled with the app; no audio is uploaded, stored, or sent to us. When you press Answer, the transcribed text of the question is sent to get an answer — the recording never is.

Does it work with Zoom, Teams and Meet?

It captures your computer's audio output, so it works with anything that plays sound. It doesn't integrate with any of them, which is why there's nothing to install on the other side.

Is it visible if I share my screen?

The overlay asks macOS to exclude it from screen capture, and it never takes keyboard focus. We've measured this working on macOS 15, but no application can promise it for every version and every conferencing tool — so test it on your own setup with the built-in self-test before relying on it.

Do I need an API key?

No. Answers come through the Claude Code CLI you already have installed, on your existing subscription.

What do you store?

Your email, your credit balance, and how long each session ran. Your résumé, interview context and session transcripts stay on your machine. Questions you're asked are contributed to the question bank that powers Learning — you can turn that off in your account settings at any time.

Windows?

Not yet. macOS 14 or later for now; a Windows version is in progress.

macOS 14 or laterSonoma and up
Apple siliconM1 and later
An 8 MB appSpeech model downloaded once, 141 MB or 514 MB
30 minutes freeNo card to start

Try it on your next interview

30 minutes of credit, no card, about a minute to set up.

Sign in with Google