OpenRewind

Rewind your Mac or PC. OpenRewind records your screen and audio, pulls the text out, and lets you search or scroll back through everything you saw and heard. 100% local. Open source. A small Swift app on the Mac, a native .NET app on Windows.

OpenRewind was called openscreen until October 2026. On first launch it moves ~/.openscreen to ~/.openrewind and keeps every recording and your menu settings. macOS treats it as a new app, so it asks for Screen Recording, Microphone and Accessibility again; you can then delete the old OpenScreen.app (and its Login Item).

The story

2022: a search engine for your life

On November 1, 2022, Rewind launched with $10M from Andreessen Horowitz, at a $75M valuation. The founders were Dan Siroker (co-founder of Optimizely) and Brett Bejcek.

The idea was simple and a little wild. Your Mac records everything on your screen and everything said in your meetings. Then you can search any word and jump back to that moment. It all ran on your own Mac, made possible by Apple Silicon. Rewind promised it would "never sell your data or do advertising."

2023: the hottest pitch on Twitter

In April 2023, Rewind raised its Series A in public. Siroker posted the pitch deck and asked investors to send offers. More than a thousand investors showed interest and 170 made offers. NEA led the round at a $350M valuation.

2024: the pivot away from the Mac

In April 2024, Rewind renamed itself Limitless and moved to a $99 AI pendant that records your conversations. The Mac app stopped getting updates.

2025: lights out

On December 5, 2025, Meta bought Limitless. The team moved to Meta's wearables group. The Rewind Mac app was shut off: on December 19, 2025, all screen and audio capture stopped. Users in the EU, UK, and some other countries had two weeks to export their data before it was deleted.

Now: OpenRewind

A tool that remembers your whole digital life should not be able to vanish when a company is sold. So we built it again, in the open:

• Local only. Your data stays in ~/.openrewind. No account. No cloud.
• Open source (MIT). Read the code. Change it. Nobody can switch it off.
• Plain files. SQLite and images. Your history is yours to keep or move.

How it was built

OpenRewind is new code. It does not contain any Rewind code. The idea and the full-screen timeline (⌥⌘T) are modeled on Rewind's.

OpenRewind is not made by or connected to Rewind AI, Limitless, or Meta.

What it does (v0.1)

• Takes a screenshot every 2 s (ScreenCaptureKit).
• Skips frames that did not change. Settings → Save Frames While Typing can save fewer frames (every 5, 10 or 30 s) while only a few words change; the default is every 2 s, and the screen you stop on is always saved.
• Reads text from the focused window's accessibility tree. Falls back to Apple Vision OCR. Never reads password fields.
• Saves a HEIC (about half the size of a JPEG; JPEG on Macs that cannot encode HEIC) and a row in SQLite with an FTS5 index.
• Can serve a local API + search page at http://127.0.0.1:3031 (off in the app until you turn it on).
• Blocked apps are blurred in every screenshot, wherever they are on screen, and their text is never saved. First launch asks what to block (password managers, money apps, private browsing windows and bank sites are ticked); change it in Settings → Blocked Apps & Sites. OPENREWIND_EXCLUDE adds more app names from the terminal.
• Records microphone (AVAudioEngine) and system audio (ScreenCaptureKit) in 30 s chunks, 16 kHz mono WAV. Silent chunks are dropped.
• Transcribes each chunk on your computer with local Whisper (whisper-cli from whisper.cpp) or NVIDIA Parakeet v3, your pick in Settings → Transcription Model (see Choosing a transcription model). Transcripts go into their own FTS5 index and show up in the same search as screen text.
• Ask AI (menu bar or tray → Ask AI…) answers questions like "what did I work on today?" with your own key: Anthropic, OpenAI, Gemini, OpenRouter (any model it offers), or any OpenAI-compatible server (Ollama, LM Studio). See Ask AI.
• Search by Meaning (off until you turn it on) finds moments by what they were about, not only the exact words. It runs on your computer. See Search by Meaning.
• Journal (menu bar or tray → Journal…): one page per day with your top apps, sites and windows, worked out on your computer, plus an optional AI recap. See Journal.

Compress Older Recordings

Settings → Compress Older Recordings keeps screenshots older than an hour as short videos instead of one picture each. Screens change little from one frame to the next, so a video takes about 5 to 10 times less disk space. It is off by default: it is new and has not been tested on many Macs and PCs yet. The storage line in the menu shows what the videos save ("Using 3.1 GB · video saves 14.2 GB").

• Each video holds up to five minutes of one screen (frames of different sizes go in separate videos). The last hour stays as pictures, so the recent timeline is exactly as before.
• The Mac uses HEVC (H.265) when it has a hardware HEVC encoder, else H.264; every Mac this app runs on can play both. Windows always uses H.264: playing HEVC on Windows needs the optional HEVC Video Extensions, and a data folder restored on a PC without them would show nothing for its older recordings. The Mac uses AVFoundation, Windows uses Media Foundation, both built into the system; nothing is downloaded and nothing leaves the computer.
• Captures wider or taller than 4096 x 2304 (ultrawide and 5K screens) are scaled down to fit, since that is the most every H.264 encoder takes. A screenshot that still fails to go into a video three times stays a picture.
• Crash-safe: a video is written to a temporary file, read back and compared with the pictures, moved into place, and the frames are pointed at it in one database transaction. Only then are the pictures deleted. Quitting or crashing at any point loses nothing; the next run cleans up.
• The timeline, search highlights, the local API (GET /frames/{id} still returns a JPEG), and Ask AI's screenshots all read frames through one loader. It seeks exactly to each frame, keeps the last few videos open (closed after 30 s unused on Windows) and the last 48 frames decoded (at most 256 MB), so scrubbing stays quick. A keyframe every 20 frames means one frame never needs more than 20 decoded to show it.
• Deleting recordings and Keep Recordings For work the same: a video whose frames are all deleted is deleted; one that loses only some of them is written again without them, soon after. The delete and the note that the video must be redone are one database transaction, and every pass also checks for videos holding more pictures than frames point at (a delete by an older copy of the app, say), so deleted pictures can't be left in a video. If a video can't be written again (no encoder any more, a damaged file), the frames it keeps are saved as pictures and the video is deleted. Until that is done the menu and doctor say how many videos still hold deleted frames.
• Video is lossy. Each frame is checked against its screenshot when the video is made, but it is not pixel for pixel the same. Writing a video again after a delete decodes it and encodes it again, which loses a little more; that happens about once per video (Keep Recordings For trims whole five-minute videos most of the time).
• Turning it off stops making new videos. Videos already made keep playing.
• openrewind video-check (Mac) or OpenRewind.exe video-check (Windows) checks that this computer can make the videos: it compresses made-up screenshots in a temporary folder, reads every frame back, deletes some and checks the rewrite, compresses a capture wider than 4096 pixels, and (on the Mac) turns a video back into pictures. shows which codec would be used, how much space the videos save, and whether the videos already made still open (after moving the data folder to another computer, say).

Search by Meaning

Keyword search finds the words you type. Search by Meaning also finds moments that are about the same thing in other words: "the flight I was booking" finds the airline page, "when we talked about hiring" finds the call where someone said "we need another engineer".

Turn it on in Settings → Search by Meaning. OpenRewind then builds an index in the background: it goes through your old recordings first and then keeps up with new ones. It works slowly while you use the computer, slower still on battery (or in Low Power Mode on a Mac), and faster when you are away. The line under the switch shows how far it has got. If the index can't be saved (for example, the disk is full), the line says Paused. OpenRewind tries again after 1 minute, then waits twice as long each time, up to 1 hour. After 5 failed tries in a row it skips that one moment. The reason is in openrewind.log.

Once it is on, the timeline's search box has a Keyword | Meaning switch. Meaning shows the closest moments first; click one to go there, as with keyword search. The browser search page and Ask AI (search_by_meaning) can use it too, and so can AI apps connected through MCP.

How it works:

• A "moment" is a stretch of one window (until the app, window or page changes, up to 10 minutes), or one 30-second audio clip. Repeated text in a stretch is counted once.
• Each moment gets a sentence vector made on your computer. The Mac uses the sentence model that comes with macOS, so nothing is downloaded. Windows downloads one model the first time (all-MiniLM-L6-v2, about 90 MB, from Hugging Face), checks it against a fixed SHA-256, and runs it with ONNX Runtime. Your recordings are never uploaded.
• The index is stored in the same database. Each moment keeps a copy of its text (up to about 1,500 characters of screen text or transcript) with its app, window title and web address (table meaning_moments), and its vector (table meaning_vectors): up to about 2 KB per moment. So the index holds screen text too, not only numbers, and is deleted with it.
• Blocked apps, sites and windows are never indexed, and are hidden from results if you block them later. As with Ask AI, audio is not filtered by app. Apps you unblock are only indexed from then on.
• Deleting recordings (by hand or with Keep Recordings For) deletes their vectors and text too, in the same step as the recordings, so nothing from them stays searchable. Turning Search by Meaning off deletes the whole index; Windows keeps the downloaded model so turning it back on downloads nothing.
• The Mac model and the Windows model are different, so their scores are not comparable. The Mac uses your system language when macOS has a sentence model for it, otherwise English.

Ask AI

Open it from the menu bar or tray (Ask AI…) or with the Ask AI button in the timeline, which opens the same chat beside the timeline.

Set it up in Settings → AI Settings…: pick a provider, paste your key and pick a model. Type any model name, or Refresh List to ask the provider which models it has, so new models work without an app update. Keys are kept in the Mac Keychain or the Windows Credential Manager, never in the settings file or the database.

Requests go straight from your computer to the provider; there is no OpenRewind server. The model never gets your whole history. It calls lookups that run on your database and send back capped, trimmed text:

Lookup	What it sends
activity_summary	Time per app, window and website, and a timeline of what was in front
search_screen	Matching moments from the screen-text search, with a short snippet
search_audio	Matching transcript lines (or all of them in a time range)
search_by_meaning	The moments closest in meaning to a question, with a short snippet. Only offered while Search by Meaning is on
get_screen_text	Screen text from a short time range, sampled and trimmed
view_screenshot	One screenshot near a moment, as a JPEG. Only offered when you turn on Let AI look at screenshots (off by default)
list_meetings	Meetings in a time range, with their summary, decisions and action items when they have notes (see Meetings)
list_starred	Moments the user starred (⌥⌘S / Ctrl+Alt+S), with what was on screen; blocked apps and sites are left out
list_promises	Things the user promised (what, for whom, due), open ones first. Only offered while You promised is on; blocked sources are left out

Each lookup shows in the chat with the arguments the AI asked for and exactly what was sent back. Times in answers link back to that moment in the timeline. Each answer shows the tokens it used and what they cost; the header shows the chat's total and AI Settings keeps a running total per provider. Prices are built in for Claude and GPT-6 models and come with OpenRouter's model list; for other models you can type the price. Lookups leave out every app, site and window on your block list, including anything saved before you blocked it (the activity summary still counts a blocked app's time, as "(blocked app)", without its name); audio transcripts are not filtered by app. Web links in an answer only open if you or a lookup supplied that address, so text on screen can't trick the AI into a link that sends your data elsewhere. With a local model (Ollama, LM Studio) nothing leaves your computer at all; a custom server reached over plain http on another network gets a warning, since your key would travel unencrypted.

Journal

Open it from the menu bar or tray (Journal…). Each page is one day: ‹ and › go to the day before or after, the date picker jumps to any day, and Today comes back.

Every page has a part worked out on your computer, with no AI and nothing sent anywhere:

• first and last activity and the total active time (a screenshot counts until the next one, at most 2 minutes)
• top apps, top websites and top window or document titles
• when audio was recorded (microphone, system audio or both)

Apps, sites and windows on your block list are never shown: a blocked app is left out of every list, and a blocked site or window keeps only its app name. Screenshots saved while a blocked window or site was in front are left out of every list too (their time still counts as active). Click a time to open that moment in the timeline.

Recap: with an AI provider set up in AI Settings, Write Recap asks your model for a short recap: what you worked on, who you talked to, and open loops to follow up. It sends that day's history with your key: app and window names, web addresses, transcripts and some screen text.

• It uses the same lookups as Ask AI, and every lookup (including the ones the model asks for) is limited to that day. What was sent lists each lookup and exactly what went to the provider.
• Blocked sites and windows are left out, and a blocked app shows only as "(blocked app)" with its time. Audio transcripts are not filtered by app, so a call heard while a blocked app was open is still sent.
• The page shows the tokens and cost. Times in the recap link to that moment in the timeline.
• Recaps are saved in the database (journal_recaps), so a day is not written (or paid for) again; Rewrite replaces it. One written before the day was over says so.
• Each recap remembers the block list it was written with. If you block something new later, older recaps are hidden (and left out of exports) until you rewrite them, since they may mention it.

Settings → Write Each Day's Recap Automatically (also a switch in the Journal window) is off by default because it spends tokens. When on, it writes yesterday's recap the first time the app is running on the next day, with the AI provider it was turned on for; after you switch providers in AI Settings it pauses until you turn it on again.

• Days with nothing recorded are skipped. A day the app is not running the day after (say, a long weekend with the computer off) gets no automatic recap; open it and use Write Recap.
• A recap you wrote by hand during the day is replaced by one of the whole day.
• If it fails (offline at login, a busy provider), it tries again 10 minutes later, at most 3 times for that day.

Deleting recordings: deleting a time range by hand (Delete Recordings → Last Hour, Today …) also deletes the recaps of the days it touches, and deleting all recordings deletes every recap. Keep … Recordings For deletes a day's recap only once nothing of that day (screen or audio) is left, so with Keep For 1 Day yesterday's recap stays while part of yesterday does.

Copy as Markdown and Save as Markdown… export the page. In the recap, times become plain text, any other link or image becomes its text and address in plain text, and HTML is escaped, so opening the file never loads anything.

Copy, export and delete

Get things out of the timeline, on the Mac and on Windows:

• Copy text from the past: drag across words on any screenshot in the timeline to select them, then ⌘C / Ctrl+C. The Mac uses Live Text (VisionKit); Windows reads the shown screenshot with Windows OCR. Both run on your computer.
• Copy All Text (⇧⌘C / Ctrl+Shift+C, or right-click the screenshot) copies all the text saved for the shown moment, from the accessibility tree or OCR, even for frames recorded long ago.
• Copy Screenshot: ⌘C / Ctrl+C with no text selected copies the whole screenshot as an image. Save Screenshot… (⌘S / Ctrl+S) saves it as a PNG.
• Export Clip… (⌘E / Ctrl+E) turns a stretch of the timeline into an MP4 video or an animated GIF. Mark where it starts with [ and where it ends with ] (the bar shows the marked range; Esc clears it), or leave it unmarked to get the 5 minutes before the shown screenshot. Pick the format and a speed (1× real time, or 2×, 5×, 10×, 30× faster). Before you start, it shows how long the clip will be, how many frames it has, its size in pixels, and roughly how big the file will be and how long making it takes. Then pick where to save it. It says Preparing… at once, then shows a bar with the frame it is on ("Frame 120 of 600"), the percentage, the time so far and about how long is left. Cancel stops it right away and leaves no file. When it is done, Show in Finder / Show in Explorer and Open take you to the file.
• The share button at the top of the timeline (or a right-click on the screenshot on Windows) has all of these, plus Copy All Text and Mark Clip Start / End Here.

Delete things from the timeline, on the Mac and on Windows (each asks first, and can't be undone):

• The trash button at the top of the timeline has all of these.
• Select a stretch: ⇧-drag / Shift-drag across the bar at the bottom (or pick Select a Stretch to Delete… from the trash button, then drag), or press [ where it starts and ] where it ends. A bar over the dock shows what is selected, with Delete…, Export Clip… and ✕. A stretch takes only what the timeline shows in it: with the app filter on, only that app's screenshots go ("Delete 312 screenshots of Slack"), and apps hidden from the timeline are never touched. With all apps shown, you choose whether the audio clips that start in it go too.
• ⌫ / Delete deletes the selected stretch, or the shown screenshot when nothing is selected.
• Delete Every Screenshot with "…" (trash button, while keyword search has words in it) deletes every screenshot that has those words, in any app and on any day: for a password or key that was on screen.
• What was made from them goes too: the screen text and its search index rows, Search by Meaning moments, stars on them, meeting titles taken from them and the journal recap of the day. Videos are written again without them (see Compress Older Recordings).
• The deleted text is not left in the database file: SQLite overwrites what a delete frees (secure_delete), the search index drops the words at once (FTS5 secure-delete; Windows' SQLCipher is too old for it, so there the index is merged after a delete, a little at a time), and the write-ahead log is emptied. The same goes for Delete Recordings in the menu.

What is left out:

• Screenshots saved while a blocked app, window or site was in front, and ones whose front app, window or site is on your block list now (even if saved before you blocked it), are left out of clips and can't be copied, saved or read for text. With the app filter on, a clip has only that app's screenshots, as the timeline shows them.
• Clips are made from the screenshots already saved, never from a new recording. Blurring of other windows happens when a screenshot is taken, using the block list as it was then. So an app you block today is still visible, unblurred, in older screenshots where another app was in front and it was showing beside it. Check a clip before you share it if you blocked something recently.
• Copies go to the system clipboard. On Windows OpenRewind marks them so they stay out of clipboard history (Win+V) and cloud clipboard sync. On the Mac, Universal Clipboard can still hand a copy to your other Apple devices nearby while Handoff is on; turn Handoff off in System Settings → General → AirDrop & Handoff to stop that. The Live Text menu on the Mac has only Copy and Select All; web search, translate and look up are taken out.

A screenshot lasts until the next one, at most a minute, so long breaks are skipped. A clip covers at most 2 hours; videos are at most 1920 pixels wide and GIFs 1280, with at most 600 pictures (fast parts drop pictures to stay under that). MP4 uses H.264 (AVAssetWriter on the Mac, the encoder built into Windows). GIFs use a small encoder in OpenRewind on both, which writes as it goes, only stores the part of each picture that changed, and stops at once on Cancel.

Meetings

OpenRewind finds your calls in what it already records; nothing extra runs during a call. A meeting starts as soon as a call is on your screen: a Google Meet meeting (not Meet's start page), Zoom's meeting window or a Zoom meeting in the browser (not Zoom's home), a Teams call or meeting window, a Webex meeting, a FaceTime call or a Slack huddle. Another window of these apps (say, Zoom or Teams with no meeting name in its title) counts only once other people can be heard (system audio) for a minute. A meeting goes on while its window or page is in front; after you switch away, your microphone keeps it going for up to 30 minutes (say, while you share slides) and other sound for up to 15. It ends after 3 minutes with none of these, and calls in the same app less than 10 minutes apart (a dropped call, a rejoin) become one. Blocked apps, sites and private windows never become meetings. Without Whisper, audio waiting to be transcribed does not count as sound.

Take notes? Within about 10 seconds of a call coming to the front, even before anyone speaks, a small popup asks "Meeting detected in Google Meet. Take notes?" in the corner of the screen. It does not take focus from the call, and it stays while you switch to other windows (until a while after the call ends). It asks only once per call. If you say Take Notes, the call is kept with its transcript from the moment you joined, even if you spend the call in other windows.

• Take Notes keeps the call in Meetings with its transcript and, if summaries are on (below), sends it for a summary when it ends.
• Not This Time leaves the call out of Meetings and Ask AI, and it is never summarized. Audio is still recorded as usual under your recording settings; delete it from Delete Recordings if you want it gone.
• Remember for Zoom answers the same way for that app from then on. Settings → Forget Remembered Meeting Choices asks again for every app.
• Settings → Ask Before Taking Meeting Notes (on by default): turn it off to take notes for every call without asking (apps you said Never for are still skipped). If you don't answer, the call is listed but not summarized by itself; you can still press Summarize.

Meetings… in the menu bar or tray lists them by day. Each one shows its transcript, built from the audio transcripts of that time: lines from your microphone are Me, lines from system audio are Others (OpenRewind does not tell people apart). Open in Timeline jumps to the start, Copy as Markdown and Save as Markdown… export the notes and transcript.

Summaries are off by default. They use your Ask AI provider and key, cost tokens, and send the call's whole transcript to that provider. The transcript is not filtered by app: anything said or played during the call is sent, even while a blocked app or site was in front. Turn on Settings → Summarize Meetings with AI to have each call you take notes for summarized when it ends (only calls that end after you turn it on), or press Summarize on one meeting. Summaries go only to the provider you agreed to; if you switch providers in AI Settings, OpenRewind asks again and pauses summaries until you answer. If Whisper is still catching up when a call ends, OpenRewind waits up to 2 hours, and makes the notes once more if more of the call is transcribed afterwards.

You get a short summary, decisions and action items, with the tokens and cost it used, as in Ask AI. Notes never contain web links (only links into your timeline), since the AI writes them from what other people said. Ask AI can read them too: "what did we decide in my 2pm call?" Deleting audio recordings deletes the meetings in that time, and they are not found again from what is left; deleting screen recordings clears the window titles they kept.

Insights

Open it from the menu bar or tray (Insights…). Switch between Day and Week, and use ‹ › (or ← →) for the day or week before or after. It shows:

• Active time: how long you used the computer, and the change from the day or week before (up to the same time while the day or week is still going).
• Apps and Websites: time in each, as bars, the top 10 and then "Other", each with the change from the day or week before.
• When you were active: a strip of the day's hours, or for a week a grid of days by hours.
• Focus sessions: the longest stretches of 10 minutes or more mostly in one app (short looks at another app, under a minute, don't break one).
• App switches: how often the app in front changed while you were using the computer.

Click an app to open the timeline showing only that app, at the first moment you used it that day or week. Click a website or a focus session to jump to it.

Everything is worked out on your computer from what was saved with each screenshot (time, app, window title, address); nothing is sent anywhere and no AI is used. The window follows Settings → Timeline Look.

Blocked apps and sites. Screenshots saved while a blocked app, window or site was in front are marked as blocked and left out entirely, as is anything that matches your block list now (by app name, window title or address). One gap: screenshots saved by an older version while a blocked site or window was in front kept only the browser's name, and ones from a helper app of a blocked app (Docker.app shows "Docker Desktop") kept the helper's name, so that older time can still show under the browser or the helper app. Block the helper's name too to hide it.

How time is counted. Time counts from one screenshot to the next. A break of more than 2 minutes (locked, asleep or paused) does not count. A screen whose text doesn't change only gets one screenshot a minute; after 5 minutes of that, the rest counts as away until something changes, unless speech was recorded at the time (a call or a video, with Record Audio on), which counts as being there. With audio recording off, a long call or video can count as away after 5 minutes. If you record from the terminal with --interval of 50 seconds or more, every screenshot looks like a still screen and active time is undercounted; keep the interval under 50 seconds. Ask AI's activity summary counts every gap up to 2 minutes without the still-screen rule, so its totals can be a little higher than Insights'.

The first launch after updating builds a database index for Insights in the background (seconds, or about a minute for a very large history on a slow disk). Until it is done, Insights still works, just more slowly.

Use with Claude and other AI apps (MCP)

Claude Desktop, Claude Code, ChatGPT, Cursor and other apps that support the Model Context Protocol can look things up in your history with the same lookups as Ask AI (the table above). Ask them "what was I reading about databases yesterday?" and they search OpenRewind on your computer.

It is off by default. Open Settings → Connect AI Apps… and click the button for your app:

Button	What it changes
Add to Claude (Claude Desktop)	the openrewind entry in claude_desktop_config.json (Mac: ~/Library/Application Support/Claude/; Windows: %APPDATA%\Claude\, or %LOCALAPPDATA%\Packages\Claude_…\LocalCache\Roaming\Claude\ for the Microsoft Store version)
Add to Claude Code	runs claude mcp add --scope user openrewind -- <this app> mcp when it finds the claude command; otherwise writes the same entry in ~/.claude.json
Add to ChatGPT	the [mcp_servers.openrewind] table in ~/.codex/config.toml (Windows: %USERPROFILE%\.codex\config.toml; CODEX_HOME if you set it), which the ChatGPT desktop app and Codex read
Add to Cursor	the openrewind entry in ~/.cursor/mcp.json

Adding an app also turns on Allow AI Apps to Read My History (MCP). Each row then says Added, offers Restart Claude (or ChatGPT, Cursor) when the app is open, since apps read their setup when they start, and Remove, which takes out exactly that entry. Restart asks first, then quits the app the normal way and opens it again; a reply it is still writing, or a message you haven't sent, may be lost. If the entry points at another copy of OpenRewind (you moved the app) or has options you added, the row says so and the button reads Update. If you turned OpenRewind off in the app ("disabled": true, or enabled = false for ChatGPT), the row says Turned off and Turn On switches it back on.

• Only that one entry changes. Everything else in the file stays as it was. The new version goes to a private temporary file next to it and replaces the file in one step, only if the app didn't change the file in the meantime (otherwise OpenRewind starts over with what the app wrote), so it is never left half-written. Claude Desktop with both a Store and a classic install has two files: both change, or neither.
• The previous version is kept as one <file>.openrewind-backup next to it, readable only by you, and Remove deletes it. If the file is a link into a folder you keep in git (dotfiles), the backup is created in that folder: add *.openrewind-backup to its .gitignore.
• It is best to quit Claude Desktop before adding: it may write its own settings back when it quits. Claude Code is set up with its own claude command when OpenRewind finds it (also when installed with nvm, Volta, asdf, fnm, Bun or Homebrew).
• A file OpenRewind can't read safely (not valid JSON, not saved as UTF-8, or ChatGPT's file already sets up openrewind some other way) is never changed: the row says what's wrong and you can set it up by hand.
• ChatGPT: OpenRewind works in Work chats in the ChatGPT desktop app (Mac and Windows) and in Codex, which start local MCP servers from config.toml. What OpenRewind returns there goes to OpenAI like anything else in the chat. Regular chats and chatgpt.com in a browser can't reach it: they only connect to MCP servers on the internet.

For other AI apps, the window also shows the lines to paste, with this copy's path filled in and a Copy button. For most apps they look like this (Mac path shown; on Windows it is the path of your OpenRewind.exe):

{
  "mcpServers": {
    "openrewind": {
      "command": "/Applications/OpenRewind.app/Contents/MacOS/OpenRewind",
      "args": ["mcp"]
    }
  }
}

Claude Code by hand, run once:

claude mcp add --scope user openrewind /Applications/OpenRewind.app/Contents/MacOS/OpenRewind mcp

ChatGPT and Codex by hand, in ~/.codex/config.toml:

[mcp_servers.openrewind]
command = '/Applications/OpenRewind.app/Contents/MacOS/OpenRewind'
args = ['mcp']

How it works:

• The AI app starts OpenRewind mcp (Windows: OpenRewind.exe mcp, or --mcp) and talks to it over stdin/stdout. It shows no icon or window, records nothing and never changes the database: it opens it read-only, so it works whether or not OpenRewind is running.
• Lookups are the same as in Ask AI and capped the same way. Blocked apps, sites and windows are left out of screen text, searches and screenshots; in the activity summary a blocked app shows only as "(blocked app)" with its time. Audio transcripts are not filtered by app, so a call or video heard while a blocked app was open can show up in search_audio. One activity summary covers at most 31 days. Screenshots are only shared when Let AI look at screenshots is on in AI Settings.
• Names you add with OPENREWIND_EXCLUDE are not applied here: the AI app starts OpenRewind with its own environment. They are still never recorded; only frames saved before you added a name can show up. Put apps on Blocked Apps & Sites to hide them everywhere.
• Each moment comes with its exact local time (like 2026-10-18T14:30:05-07:00), which the AI can pass back to read more around it.
• While the switch is off, the AI app still lists OpenRewind, but every lookup answers that you need to turn it on. Turning it off takes effect on the next lookup.
• What the AI app reads goes to that app's AI provider, like anything else you ask it.
• Text on your screen (web pages, emails, messages) reaches the AI app as data. Unlike Ask AI, OpenRewind cannot stop an AI app from opening a link or running a command that such text asks for, and many AI apps (Claude Code, Cursor) can do both. OpenRewind tells the AI to treat it as data only, but keep an eye on what it does with your history.

Choosing a transcription model

Settings → Transcription Model picks what turns recorded speech into text. Every choice runs on your computer; audio never leaves it.

Model	Download	Languages	Notes
Whisper (Automatic) (default)	none	99	Uses the ggml Whisper model already on the computer (see Whisper). Needs whisper-cli.
Whisper Base	148 MB	99	Fast, less accurate. Needs whisper-cli.
Whisper Small	488 MB	99	A good balance. Needs whisper-cli.
Whisper Large v3 Turbo	574 MB	99	The most accurate Whisper, and the slowest. Needs whisper-cli.
NVIDIA Parakeet v3	670 MB	25 European	Very accurate, with punctuation and capitals, and fast on a CPU. Needs nothing else.

• Picking a model that is not on the computer yet shows its size and downloads it, with a progress bar. Files come from a fixed commit on Hugging Face and are checked against their SHA-256 before they are used (again the first time they are used after each start of the app). A download that stops, or gets no data for 60 seconds, carries on where it left off with Try Again. They are saved in ~/.openrewind/models: Whisper files where Whisper already looks, under their own names (ggml-small.openrewind.bin, so a ggml-small.bin you put there yourself is never replaced); Parakeet in models/parakeet-tdt-0.6b-v3-int8. Delete a file to free the space.
• The new model is used for audio recorded from then on. Re-transcribe Recent Audio (last hour, day or week) runs older clips through it too. It says how many clips that is and about how long it keeps the processor busy; new recordings still go first. The new text replaces the old for good (the old text stays searchable only until each clip is redone). Clips where no speech was found are left as they are, and meeting notes already made are kept: re-transcribing does not make them again or send the transcript to your AI provider again.
• The menu shows the model in use, when it last finished a clip and how long that took, and how many clips are waiting.
• Each transcript line records which model wrote it (audio_transcriptions.engine: whisper.cpp:ggml-small, parakeet-tdt-0.6b-v3, …). Search, Ask AI, the timeline and MCP read every line the same way.
• Parakeet languages: Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish and Ukrainian. It finds the language by itself; the --language option only applies to Whisper. For other languages pick a Whisper model: when --language is set to one Parakeet does not know, picking Parakeet (or re-transcribing with it) warns first.
• Parakeet runs on ONNX Runtime (MIT), built into the app (its notice is in THIRD_PARTY_NOTICES.txt, which ships with both apps). It holds about 1 GB of memory while it is loaded and lets go of it after 5 minutes with no audio to do. While audio recording is on, a clip arrives every 30 seconds, so it stays loaded the whole time; turn audio off, or pick Whisper, to get that memory back. Long audio (ingest-audio of a talk, or a long --chunk) goes through the model in pieces of at most 30 seconds, so memory stays the same however long it is. On a 4-core Linux test machine (the Windows code) a 30 s clip took 3.4–3.6 s, about 9 times faster than real time, and 5 minutes took 25 s.

NVIDIA Parakeet TDT 0.6B v3 is made by NVIDIA and licensed CC-BY-4.0. OpenRewind downloads the version converted to ONNX and quantized to int8 by the sherpa-onnx project (Apache-2.0), and uses it as it is. Whisper models are OpenAI's Whisper (MIT) in the ggml format of whisper.cpp (MIT).

Mac app

scripts/build-app.sh

This makes dist/OpenRewind.app and dist/openrewind-<version>.dmg. Copy the app to /Applications and open it, or run scripts/install.sh, which builds, copies to /Applications and restarts the app in one step. It lives in the menu bar:

• Record Screen and Record Audio switch on and off on their own (remembered).
• Pause Recording: for 30 minutes, 1 hour, 1 day, or until you resume. Both screen and audio stop; the menu shows Resume Recording and when recording comes back on (also after a restart). The switches still pick what comes back on.
• Open Timeline (or ⌥⌘T from anywhere) shows a full-screen overlay. Scroll or use ← → to move through time, ⇧ for 1 minute, ↑ ↓ for 10 minutes, drag or click the bar, space plays the audio at that moment (the picture follows along), + / − zoom, N jumps to now, Esc closes.
• O (or the button at the bottom right of the screenshot) goes back to that moment: it reopens the browser tab (Safari, Chrome, Arc, Dia, Edge, Brave, Firefox), the document the app had open (Pages, Preview, Xcode…), or brings the app's window with that title to the front.
• Star moments: ⌥⌘S from anywhere (or Star This Moment in the menu) stars what is on screen now; on the timeline S or the ☆ button stars the shown moment (again to unstar). Stars show as small yellow stars on the bar, and the Starred button next to search lists them (search words then only match starred moments). Only the time is saved; deleting recordings deletes their stars too. Ask AI and AI apps can read them with list_starred.
• App filter: the All apps button next to the search box (or A for the app in the shown screenshot) limits the timeline to the moments when one app was in front. Time spent in other apps shrinks to breaks on the bar, ← → skip straight between that app's screenshots, and search only finds that app's screen text (and speech heard while it was in front).
• Copy and export: drag across text on a screenshot and ⌘C copies it; with nothing selected ⌘C copies the screenshot, ⌘S saves it as a PNG, [ and ] mark a clip and ⌘E exports it as MP4 or GIF. See Copy, export and delete.
• Search… opens the timeline with the cursor in its search box, or the web search page when the local search server is on.
• Insights… shows where your time went today or this week (see ).

Put it online for others

Downloads open without a warning only when the app is signed with a Developer ID Application certificate and notarized by Apple:

1. Create the certificate at developer.apple.com → Certificates (needs the paid program).
2. Save notary login once:

xcrun notarytool store-credentials openrewind --apple-id [email protected] --team-id TEAMID

3. Build, sign, notarize, staple:

NOTARY_PROFILE=openrewind scripts/build-app.sh

Upload the .dmg (GitHub Releases, your site, etc.). Without these steps the app still works, but other people must right-click → Open.

Updates

Both apps ask GitHub for the latest release every 6 hours and when you open the timeline. A newer version shows a banner at the top of the timeline (✕ hides it until the next version) and a row in the menu. Settings → Download Updates Automatically (on by default) fetches it in the background; nothing is installed until you click Update / Restart to Update, which swaps the app in place and reopens it. Settings → Check for Updates… checks right away. Releases come from the Release workflow (git tag v0.3.0 && git push origin v0.3.0); the check needs the repository to be public. Screen Recording, Accessibility and Microphone stay on across updates because every build is signed with the same certificate (below).

Signing

macOS remembers Screen Recording, Accessibility and Microphone for the certificate an app is signed with. An ad-hoc signature is new on every build, so each update used to lose them. Now every build is signed with one long-lived certificate:

• Make it once: scripts/make-signing-cert.sh writes a self-signed "OpenRewind Code Signing" certificate to signing/ (valid 20 years). Keep the .p12 and password safe and don't commit them: anything signed with that key gets the permissions you gave OpenRewind.

• GitHub Actions: the real certificate is only for releases. It lives in a release environment that only main and v* tags can use, so a pushed branch or a PR can't reach it even by editing a workflow. Branch and PR builds sign with a throwaway certificate. Set it up once:

  1. Settings → Environments → New environment, name it release.
  2. Deployment branches and tags → Selected branches and tags → add the branch main and the tag pattern v*. Optionally tick Required reviewers and add yourself, so each release waits for your click.
  3. In that environment, Add environment secret: MACOS_CERTIFICATE_P12 (the contents of openrewind-signing.p12.b64) and MACOS_CERTIFICATE_PASSWORD (the contents of password.txt).
  4. Settings → Secrets and variables → Actions: delete any repository secrets with those names (and the AZURE_* ones, after adding them to the environment). Repository secrets are open to every branch.
  5. Settings → Rules → Rulesets → New tag ruleset for v*, restricting creation to you, so nobody else can push a tag that opens the environment.

  CI imports the key as non-extractable and deletes the keychain after signing. The Release workflow refuses to build the Mac app without the secrets.

• Your Mac: double-click openrewind-signing.p12 and enter the password. scripts/build-app.sh picks it up (after a Developer ID certificate, before Apple Development or ad-hoc); allow codesign to use the key when asked.

• Developer ID later: export the Developer ID Application certificate as a .p12 and put it in the same two secrets. That also removes the "unidentified developer" warning (with notarizing).

Changing the certificate (ad-hoc → self-signed, or self-signed → Developer ID) means turning the permissions on one more time; after that, updates keep them.

Windows app

Windows 10 (2004+) and 11, x64 and ARM64. Same timeline, search page, API and data format as the Mac app; the timeline opens with Ctrl+Alt+T. Download, setup and build steps are in windows/README.md:

scripts\build-windows.ps1

Run from the terminal

swift build -c release

.build/release/openrewind doctor

.build/release/openrewind

Then open http://127.0.0.1:3031.

Whisper

Needs whisper-cli (brew install whisper-cpp) and a ggml model. Settings → Transcription Model downloads one (Whisper Base, Small or Large v3 Turbo), or OpenRewind looks for a model in this order:

1. --whisper-model PATH or $OPENREWIND_WHISPER_MODEL
2. The Whisper size picked in Settings, once downloaded
3. ~/.openrewind/models/ggml-*.bin
4. VoiceInk's models folder (already on this Mac: ggml-large-v3-turbo-q5_0.bin)

NVIDIA Parakeet v3 needs neither whisper-cli nor a ggml file: pick it in Settings and it downloads.

Test it without a mic:

.build/release/openrewind ingest-audio /opt/homebrew/share/whisper-cpp/jfk.wav

macOS asks for Screen Recording (screen + system audio) and Microphone permission for the app that runs the command (for example Terminal). Grant it, then run again. Accessibility permission is optional; it gives cleaner text than OCR.

API

On when you run record or serve from the terminal, or after Settings → Local Search Server in the app.

Nothing the block list covers comes back: screenshots of a blocked app, a blocked site, a private window or a window or address with a blocked word are left out of /search (also mode=meaning), and /frames/{id} answers 404 for them, the same rule as Ask AI and AI apps. That includes screenshots saved before you blocked them, when the blocked app, site or window was the one in front. An older screenshot of another app can still show a blocked window behind it, and its text can include text from that window. /apps leaves out blocked apps. Audio transcripts are not filtered by app (a clip has no app). The total in pagination counts only what can be shown, on every page.

If a search would have to skip more than 20,000 hidden screenshots, it stops there: pagination then has "truncated": true, and total counts only what comes before that point.

Route	What
GET /health	Status, frame count, permissions
GET /search?q=&app_name=&window_name=&start_time=&end_time=&limit=&offset=&max_content_length=	Full-text search. Times: ISO, 2h ago, today, yesterday, YYYY-MM-DD
GET /frames/{id}	Snapshot as JPEG (also for frames kept in a video)
GET /apps?start_time=	Apps with frame counts
GET /audio/{chunk_id}	Audio chunk WAV

content_type: all (screen + audio), ocr (screen only), audio.

mode=meaning searches by meaning instead of words (only while Search by Meaning is on, else 400; /health says "search_by_meaning": true). Results come best first, in the same shape plus a score, and leave out blocked apps and sites. A screen result is a moment: timestamp is where it starts and text is the moment's text.

Response: {"data":[{"type":"OCR","content":{"frame_id","text","timestamp","app_name","window_name","file_path","text_source"}},{"type":"Audio","content":{"chunk_id","transcription","timestamp","start_time","end_time","device_name","device_type"}}],"pagination":{"limit","offset","total"}} (plus "truncated": true when a search stopped early, see above)

Encryption and App Lock

Both are off by default. Find them in Settings → Encryption & Lock, on the Mac and on Windows.

Encrypt Recordings encrypts new screenshots, audio clips and compressed videos (Compress Older Recordings) with AES-256-GCM, and the database with SQLCipher. The database holds everything else OpenRewind keeps: the text of each screenshot, transcripts, the Journal's recaps, Meetings and their notes, and the Search by Meaning index, so those are encrypted too. What is already recorded is encrypted in the background; quitting halfway is fine, it carries on at the next launch. Turning it off decrypts everything back the same way. The menu shows how far it has got. The database is rewritten in one go first: recording waits until that is done (a few seconds, up to a minute or two for a very large history), and it needs free disk space for a second copy of the database.

• What it protects against: someone who copies your disk, a backup or the ~/.openrewind folder, and other user accounts on the same computer. Without your login they see only encrypted files.
• What it does not protect against: apps running as you. OpenRewind has to record in the background without asking for a password, so the key is kept where your login unlocks it: the login Keychain on the Mac, DPAPI for your Windows account on Windows (~/.openrewind/encryption-key.dpapi). Anything that can run as you can ask for it too. The login Keychain is a file (~/Library/Keychains/login.keychain-db) protected by your login password; it is not tied to this Mac, so a copy of it plus your password opens the key.
• What was there before: backups, Time Machine snapshots and cloud copies made before you turned it on still hold the plain files, and the old plain files are deleted, not wiped, so they can stay on the disk until it reuses the space. Turn it on before you record anything sensitive, or delete the older backups.
• The log (~/.openrewind/openrewind.log) is a plain file. With encryption on it no longer names the app of each saved frame or meeting, or a meeting's times; earlier lines (from before it was on) still do.
• Recovery key: Export Recovery Key… saves the key as a text file (72 characters in 9 groups) that only you can read. OpenRewind offers to save it when you turn encryption on. If the Keychain or your Windows account is reset, or you move your recordings to another computer, OpenRewind asks for it at launch and checks it against your recordings. If you lose the key and the Keychain or Windows account, the recordings are lost for good. Keep it somewhere safe, like a password manager. The same key works in the Mac and the Windows app. OpenRewind never replaces a stored key it cannot read, and keeps a copy of the old one when a recovery key replaces it.
• Speech to text: each clip is decrypted in memory. Parakeet works from memory; whisper-cli needs a file, so it gets the clip as a WAV file in your temp folder that only you can read, for the few seconds it takes, deleted right after.
• Compressed videos: the video player reads files, so a video being shown (timeline, search, clip export, AI screenshots) is decrypted to a file in your temp folder that only you can read, while it is in use. It is closed and deleted after 30 seconds unused, when the timeline closes or the app locks or quits, and after each AI-app (MCP) request. A new video is made in the temp folder and saved into data/video encrypted; no plain video is written under data/. Copies left by a crash are removed when the app, record, serve or mcp next starts, and on each pass of Compress Older Recordings.

Require Touch ID or Password to Open Timeline (Mac) / Require Windows Hello to Open Timeline (Windows) asks before showing the timeline, search, Ask AI, Journal, Meetings, Insights, Blocked Apps & Sites, the data folder and setup, before opening a moment from an Ask AI answer, and before deleting recordings or changing how long they are kept. Recording keeps going while locked. An unlock lasts until the screen locks or the computer sleeps, or until Lock Again After has passed with none of those windows in use (5 minutes, 15 minutes (default), 1 hour, or only when the screen locks); then they close. While locked, the Local Search Server answers 423 Locked for search, screenshots and audio, and /health says only that it is running. AI apps (MCP) follow the lock too: with it on, they get your history only while OpenRewind is running and unlocked (within Lock Again After of its last use), and otherwise get a message saying it is locked. Turning AI apps on or off and Connect AI Apps… ask for the unlock first. Windows Hello needs at least a PIN; without it (or after it is removed) your Windows password works instead, so you can always turn the lock off. A Mac needs a login password.

The lock is against someone using your computer while you are away from it, not against programs running as you: those could turn it off in the settings file (settings.json on Windows, defaults on the Mac) or run OpenRewind serve from a terminal. Encrypt Recordings is what protects the files.

Data

~/.openrewind/db.sqlite, ~/.openrewind/data/<date>/*.heic (*.jpg on Windows and older recordings), ~/.openrewind/data/video/<date>/*.mp4 (with Compress Older Recordings on), and ~/.openrewind/data/audio/<date>/*.wav. In the API, a frame kept in a video has the video as file_path and its frame number in it as offset_index. With Encrypt Recordings on, these files keep their names but are encrypted (see Encryption and App Lock). Delete the folder to wipe everything, or from the terminal:

.build/release/openrewind delete older:week

(15m, 1h, today, all, older:day|week|month|year; add --screen-only or --audio-only.)

Not yet (Rewind had these)

• Speaker identification (who said what)
• Meeting summaries (Ask AI and the Journal recap summarize the day's transcripts, but not each meeting on its own)
• UI input events (clicks, keys)
• Browser URL per frame
• Plugins

License

MIT. Use it, change it, share it. What the apps include from others, and under which licenses, is in THIRD_PARTY_NOTICES.txt.