CamScribe is one application, not a recorder plus an editor plus a caption tool. The take, the timeline, the cards, the captions and the export are one document with one undo history.
Every claim on this page is read out of the product's own implementation status document. CamScribe is in development: all of this is built and tested, and none of it has yet been measured across a range of Mac hardware.
Four tracks, four writers, four files. Media is written as fragmented chunks that rotate every 300 seconds, and a chunk is listed in the index before it closes, so a crash leaves recovery an explicit pointer at the incomplete tail. The index is checkpointed every 2 seconds.
Recovery reassembles by copying sample data verbatim rather than re-encoding it, so a recovered recording is not a generation older than the one you made.
The mode decides which tracks exist and what the canvas is composed from. Everything after that works the same way.
The screen with the presenter beside it or over it. The camera plate has its own shape, position, corners, shadow and background, and it can be removed from the finished video without touching the narration.
No camera track at all. The same canvas, the same zooms, the same cards, the same captions.
The camera as the subject rather than the aside, for a piece to camera with no screen behind it.
Each mode composes onto any of the four canvas shapes, and the stage template gives every shape its own starting camera size, camera position and padding. A tall source, such as an upright window or a region you drew taller than it is wide, gets the presenter in the column beside it, or standing under it on 9:16 where there is no column.
Talking points are written before the take and shown during it, in all three recording modes. Every control lives on the recorder panel and on a global shortcut, so you are never reaching for a window while recording.
The list is content rather than a preference, so it clears when you set up the next recording. Your look, your pace and your saved libraries carry over.
The framing guide answers the question a self view does not: not "where am I" but "where should I be". After the take, the editor can set the camera centre from the face it actually finds rather than from where you guessed it was.
None of this is a black box you have to trust. Every on-device model is checked against a signed manifest before it is loaded, and every one has a deterministic rule behind it that takes over when the model is missing, when verification fails, when the Mac is thermally stressed, when Low Power Mode is on, or when you have simply switched it off. That rule is a tested path rather than dead code: CamScribe records, edits and exports correctly with no models installed at all.
Seven track kinds over three layers. Clips are sparse and placed absolutely, so a gap is the absence of a clip rather than a spacer, and a rounding difference on one track can never shift everything after it relative to another track. Positions are integers on a 90 kHz timebase and clip speed is an exact reduced fraction, so drift does not accumulate with length. A looped clip tiles a passage rather than duplicating it.
56 reversible commands, each computing its own exact inverse from the live document, with a 256 entry history and gesture coalescing over a 320 millisecond window. Chapters, titles and captions are siblings rather than special cases, and captions sit on their own lane so hiding subtitles does not hide your lower thirds.
Zoom, blur, highlight, mask and volume are one automation span with one envelope, one set of handles and one mental model. Learn to place one and you can place all five.
Each parameter has a documented neutral value and documented bounds, and the shape and feather belong to the ones that read them. A span is either bounded or it covers the whole timeline, and its range is re-derived every time the document is sanitized.
Select the region, record it again, and it is spliced back in place. The editor does not need to be closed to do it.
A title over a shot competes with the shot. A card between shots does not. CamScribe treats a card as an interstitial: during the card there is no recording underneath it.
The same icon set the cards draw from, over the picture, for the range you want it. An agent can place and remove one too.
On macOS 26 and later, transcription uses the system's newer speech analyser, which takes the whole file in one call and carries a time range per word. Below that it falls back to the older recogniser, which is a tested path rather than dead code.
The difference was measured rather than assumed, and it was large: the older recogniser answers with one utterance per pass and says nothing about the rest, so a whole recording handed to it in one call can come back as its last sentence alone with no error and no truncation flag. That is why the fallback path cuts a recording into passes at its own silences rather than at a fixed length.
Search across every recording you have made, over what was said and over the window titles that were in front of you at the time.
The index lives in the recording’s own folder. There is no global database, and deleting a recording deletes its index with it.
Email addresses, phone numbers, postal addresses, web addresses, card numbers, long digit runs, credential shaped tokens and personal names are stripped before anything is written.
A result shows the moment, the second it happened, which channel found it and what kind of evidence it is, so a window title is never read as something somebody said.
With no sentence encoder present it degrades to exact term matching, still ranks, and says so rather than pretending otherwise.
There is no search tool and no search resource on the MCP surface, because a search result that returned passage text would undo the decision to report caption lengths rather than caption words.
Nothing in CamScribe performs text recognition over frames. The window context it does index comes from the window list the recorder already walks.
Preview and export are one Core Image pipeline, compositing in linear light, held to being one by a test that renders both and compares them.
It is built, it is tested, and it has never met your hardware. If you record on a Mac and you would rather tell us what is wrong than wait for it to be finished, ask Olib AI for a build.