It hears the question.
You give the answer.
takeovr listens to your call, reads your screen, and streams an answer in under two seconds — to an overlay the screen share cannot see, because the operating system never draws it into the frame.
No account needed to try it · Free tier is not a trial
class TokenBucket:
def __init__(self, rate, cap):
self.rate = rate
self.cap = cap
self.tokens = cap
self.ts = time.monotonic()
def take(self, n=1):
now = time.monotonic()
self.tokens = min(
self.cap,
self.tokens + (now - self.ts) * self.rate,
)
self.ts = now
if self.tokens >= n:
self.tokens -= n
return True
return Falseclass TokenBucket:
def __init__(self, rate, cap):
self.rate = rate
self.cap = cap
self.tokens = cap
self.ts = time.monotonic()
def take(self, n=1):
now = time.monotonic()
self.tokens = min(
self.cap,
self.tokens + (now - self.ts) * self.rate,
)
self.ts = now
if self.tokens >= n:
self.tokens -= n
return True
return FalseToggle it. The overlay never leaves the left pane.
Four things happen between the question
and the answer.
It hears both sides
System audio is captured through a loopback device and your microphone separately, so the speaker labels come from the source rather than from guessing at voices. Deepgram streams the transcript back in under a second.
It notices a question
When a line from them ends on a question mark, an interrogative opener, or a phrase like "walk me through", the answer fires on its own. No hotkey, no tell. Debounced so trailing words do not trigger it twice.
It reads your screen
A single frame — never a video stream — is downscaled, fingerprinted, and read by the OS text engine on-device. Identical frames are dropped entirely, so a static screen costs image tokens exactly once.
It answers where only you can see
One lead line, then at most three bullets, streaming token by token into a window the compositor omits from every capture. Skimmable at a glance, because you are mid-sentence.
Not a trick.
A documented OS facility.
takeovr calls setContentProtection(true), which sets the window’s display affinity. The compositor then leaves those pixels out of any capture buffer — Zoom, Teams, Meet, Discord, OBS, Snipping Tool, PrintScreen. It is the same facility password managers and DRM video players use. Nothing is injected into the conferencing app, and nothing is hooked.
A phone camera pointed at your monitor, or a hardware capture card between the GPU and the display. Display affinity is enforced by the compositor, so anything downstream of it still sees the pixels.
WDA_EXCLUDEFROMCAPTUREOmitted entirely. Participants see straight through it.
NSWindowSharingNoneOmitted entirely, same as above.
WDA_MONITORRenders as a black rectangle in captures rather than disappearing.
—No content protection exists. The overlay will appear in shares.
The overlay is additionally frameless, kept off the taskbar and out of Alt+Tab, click-through by default so it never intercepts a click meant for the call, shown without stealing focus, and excluded from its own screenshots — the model sees exactly what the other participants see.
You pay the model bill, so we made it small.
Nothing here streams continuous video. Frames go out only on a trigger, and then three more things happen before any of it becomes tokens. Roughly $0.02–0.04 per answer on Opus 5; severalfold less on Gemini Flash or Grok.
Unchanged frames cost nothing
Every capture is fingerprinted with a 64-bit perceptual hash. If the screen has not visibly changed, the image is dropped and only the cached text is sent.
The text is read on-device
Windows.Media.Ocr and Apple Vision run as a warm resident worker and answer in about 40ms. Text costs a fraction of the same pixels, and identifiers and stack traces survive verbatim instead of being guessed from glyphs.
Fewer pixels leave at all
Long edge capped at 1280 and encoded JPEG q72 rather than PNG. Measured on a real desktop: 74 KB against 266 KB — 72% smaller, with no loss of legibility for IDE text.
Nothing you press is on screen.
Global hotkeys, so they work while the call window has focus. Deliberately avoids Ctrl+Enter and its neighbours, because a global shortcut is taken system-wide and you still need those inside Slack and your editor.
The conversation is free. The screen is not.
Free is the whole conversational product — it listens, transcribes, answers, and stays hidden. Pro is what happens when it can also see.
Free
The whole conversation, none of the pixels.
forever
- Live transcription of both sides
- Streaming answers from any provider
- Auto-answer on questions it hears
- Invisible to screen shares and recordings
- All twelve hotkeys
- No screen context
- No OCR
- 12-line transcript window
Pro
It sees what you see.
billed $96 yearly
- Everything in Free
- Screen context — answers about what is on screen
- On-device OCR so code and numbers survive verbatim
- Custom profile prepended to every prompt
- 200-line transcript window
- Custom endpoints — point it at your own proxy
Team
Pro, times everyone.
billed $96/seat yearly
3 seat minimum · $288 per year
- Everything in Pro, on every seat
- Three seats minimum
- One licence, many machines
- Seat count you can raise at any time
- Consolidated invoice
- Priority support
Every tier is bring-your-own-key: you add your own Anthropic, OpenAI, Google or xAI key and pay them directly for usage. takeovr never sees your calls, and there is no per-minute charge from us.
The awkward ones, answered.
Can the other side detect it?
Not in the video. The window is excluded by the compositor, so the pixels are never in the outgoing frame — there is nothing to notice, no flicker and no artefact. What it does not do is hide the process: software that enumerates running applications on your machine, which some proctoring and endpoint-management agents do, can see that takeovr is installed and running. If something with that level of access is watching, treat it as visible.
Is it allowed?
That depends entirely on where you are using it, and it is your call to make. Many employers restrict recording or assistance on calls, and essentially every proctored exam and certification prohibits it outright — using it there is misconduct regardless of whether it is detected. Plenty of other uses are uncontroversial: your own meetings, your own customer calls, your own notes. We are not in a position to police the difference, so we are telling you plainly instead.
What does it cost to run?
You bring your own API key and pay the provider directly. A typical answer is 2–4k input tokens plus about 1.1k more if a screenshot is attached and not skipped, and 100–200 out — roughly $0.02–0.04 on Opus 5, and severalfold less on Gemini Flash or Grok. A subscription from us is for the app, not the inference.
Which models can it use?
Claude, GPT, Gemini and Grok. Add a key for at least one; the header dropdown switches provider and model live, mid-call. The model list is populated from each provider’s own endpoint, so something released tomorrow is selectable without waiting for us to ship an update.
Does it work on a Mac?
Partly, and we would rather say so up front. The invisibility works — macOS has the same facility. But Electron’s loopback audio device is Windows-only, so a Mac install transcribes your microphone and not the other side of the call. Getting the full experience on macOS means installing a virtual audio device such as BlackHole and routing your call app through it. Windows is the tested, first-class platform.
Where do my calls go?
To whichever model provider’s key you entered, and nowhere else. There is no takeovr server in the path — your key is stored encrypted by the OS (DPAPI on Windows, Keychain on macOS) and the app talks to the provider directly. We could not read your transcripts if we wanted to.
What happens if I cancel?
Your licence key keeps working until a few days after the period you have already paid for, then stops. The app drops back to the free tier rather than locking you out: it goes on listening, transcribing and answering, and just stops attaching screenshots. Nothing is deleted and no call is interrupted.
Why does Windows warn me when I install it?
The builds are not code-signed yet, so SmartScreen shows an “unrecognised app” notice — click More info → Run anyway. It is a distribution problem rather than a safety one, and it disappears once enough people install a given build. The installer is per-user, so it needs no admin rights and raises no UAC prompt.
Next call is in nine minutes.
Installs per-user, so there is no admin prompt. Add one API key and it is listening.