# mcap A browser for a library of screenshots and photos: every `.mcap` file in a folder (and its folders) as a grid of thumbnails, searchable by the text in the images (OCR), by what they show (labels), by their window title or file name, URL, app or camera and annotations, and by meaning ("a circuit board"). Written in [Jai](https://jai.community) with [Dear ImGui](https://github.com/ocornut/imgui), for Windows (macOS and Linux build, but OCR and labels are tested on Windows with an NVIDIA GPU). The `.mcap` files come from the **Import** window, which takes any folder of images (PNG, JPEG, HEIC, AVIF, WebP, GIF, BMP, TIFF, camera RAW), videos (MP4, MOV, MKV, WebM, AVI, WMV, MPEG) and Snagit `.snagx` captures, or from `snagit-converter` (`D:\McapIndex\SnagitConverter`); `modules/Capture_File/README.md` describes the format. ## Getting started 1. **Fetch what is not in git** (each download's SHA-256 is checked): ``` cd modules/onnxruntime && jai fetch.jai # onnxruntime 1.30.0, built for CUDA 13 cd modules/OCR && jai fetch_models.jai # the OCR models (~196 MB; `- mobile` for the light one only) cd modules/Labels && jai fetch_models.jai # the label model, SigLIP 2 (~690 MB) ``` 2. **For the GPU** (optional; without it everything runs on the CPU, several times slower): an NVIDIA driver, the CUDA Toolkit 13.x, and cuDNN 9.x for CUDA 13. cuDNN installed in its usual place (`C:\Program Files\NVIDIA\CUDNN`) is found without changing PATH. The OCR window says what, if anything, is missing. 3. **Build and run:** `jai first.jai - release`, then `bin\mcap.exe`. The build copies onnxruntime and the models beside the program. 4. **The library** is `D:\McapIndex\mcap`, a folder for each place imported from (`SnagitLibraryPersonal`, `SnagitLibraryWork`, ...); `-library FOLDER` shows another folder. **Import** (button above the grid) brings in more; see [Importing](#importing). 5. **Read the library:** open the **OCR** window (button above the grid), check **Labels too**, and press **Scan the library**. Every capture is read and labelled once and the results are stored in its file, so this is needed only for new captures afterwards. On an RTX 5080 laptop, 8,600 captures took about 13 minutes with the mobile OCR model; labels add about 18 ms a capture, and the server OCR model is several times slower. ## The main window ``` [ search box .......................... ] 8609 captures Size [---o---] OCR Stats Labels Refresh [ thin progress bar while a scan runs ] +-------+ +-------+ +-------+ +-------+ | thumb | | thumb | | thumb | | thumb | newest first +-------+ +-------+ +-------+ +-------+ ``` - **The grid** shows every capture, newest first by capture date (by name while the catalog, what search needs from every file, is still loading: a line beside the count says so). **Size** sets the thumbnails' size. Click a thumbnail to open it in the Capture window. - **Delete:** hover a thumbnail and click the trash button in its corner. You are asked whether to move the file to the Recycle Bin or to delete it permanently, overwriting it with random bytes first; "Don't ask again" keeps the choice until the app restarts. Only the `.mcap` goes: the `.snagx` it was converted from is untouched. - **Refresh** looks at the folder again, after files were added or deleted outside the app. - **OCR**, **Stats** and **Labels** open their windows (below). - A **thin bar** under the search box shows a running scan; hover it for the details. ### The Capture window One capture at full size, with **Left** and **Right** stepping through the captures shown in the grid (in its order, so through a search's matches). - **Fit:** Window, Width, Height, or Off with the **Zoom** slider (and **100%**). - **OCR** reads this capture now, and labels it when the label model is up (**Read again** when it has been read); **Boxes** shows the text lines found, outlined over the image (turned with the text for turned lines). Hover a box for its text and confidence. A capture read by several OCR models has a menu to choose whose lines show. - While searching, the lines that match are highlighted, with **Matches: N** and **Previous** / **Next** to scroll to each. - **Labels:** the capture's five best labels with their scores (dim when they come from an older label list). ### The OCR window - **Model:** PP-OCRv5 mobile (light, about 85 ms a capture on the GPU) or server (more accurate, about 0.4 s). The choice is kept in the settings file (`model` under `[ocr]`). - **Running on:** the GPU or the CPU, and if the CPU, what to install for the GPU. - **Scan the library** reads every capture the model has not read yet and, with **Labels too**, labels every capture not yet labelled with the current list. While a search is active the button reads **Scan the captures shown** and covers only those. Each capture is stored as soon as it is read, so **Pause**, **Cancel** and quitting lose nothing; the next scan picks up where it stopped. Progress shows captures done, the rate, time elapsed and time left. - **Erase OCR readings from the library** (for testing) takes every OCR reading out of every file, after asking. A scan reads them again. - Below, the text of the capture open in the Capture window, line by line. ### The Labels window - The label model's status, and how many captures are labelled, and with the current list. - **The label list:** `labels.config` in the app's data folder (`%APPDATA%\mcap`), with **Open** to edit it. Each line is a name and a prompt, the phrase the model compares with the image: ``` label: "Schematic" "an electronic circuit schematic diagram" label: "Code editor" "a screenshot of a code editor showing source code" ``` A fuller description scores far better than a single word. Saved changes are picked up while the app runs; the next scan relabels every capture from the image vector stored in its file, without reading the images again, so changing the list is cheap. - **Erase labels and vectors from the library** (for testing), after asking. ### The Segmentation window **Segments** (above the grid) finds what is where in images: people, animals and things, each one apart, and stuff such as sky, wall and grass (Mask2Former, `modules/Segment`, 133 COCO classes). It works only in the library folders you choose, and never on videos. - **Library folders:** a table of every folder in the library: - **Segment:** whether scans read that folder's images. It starts off for every folder, so adding a large archive never starts a long GPU job by itself. - **Segmented:** how many of its images hold segments. - **Erase:** takes that folder's segments out of its files, after asking. The choice is kept in the library (`.library-folders`), not in the app's settings, so it goes where the library goes. - **Segment the folders set to segment** reads every image in them not yet segmented. Each is stored as it is done, so a scan can stop and start again. It runs on the OCR thread, with the same progress bar, Pause and Cancel. The model loads when the first scan starts and takes about 2.4 GB of GPU memory; on an RTX 5080 laptop a scan does about 17 images a second. - **Importing** into a folder set to segment segments the new images once they are in. The Import window's "Segment the images in this folder" checkbox is the same setting. - **Storage:** each image's segments are stored in its file: a short list (class, score, area, box) for search, and a mask of about 5 KB for drawing. See `modules/Capture_File/README.md`. ### The Stats window What OCR found in the captures shown (the whole library, or a search's matches), by the chosen OCR model: captures with text, lines, words, characters, confidence, words per capture with a breakdown, a table by year, the captures with the most words (click to open), each capture's best label counted (click one to search for it), and the commonest words (click one to search for it). When any of them are segmented, it also shows **What is in the images**: how many images hold each class (person, sky, dog ...). Click one to search `has:` for it. ## Importing **Import** (above the grid) brings a folder of screenshots, photos and videos into the library: 1. **Import from:** the folder, typed or chosen with **Browse...**. Everything under it is looked at, its folders too (not hidden ones, such as `.thumbnails`). 2. **Into library folder:** the folder in the library it goes into, such as `SnagitLibraryWork` or `Phone`. It starts as the name of the folder imported from, or the library folder that folder went into last time, so importing the same place again adds to the same folder. 3. **Import.** The progress, the time left, and Pause and Cancel show while it runs; new captures appear in the grid as they are written. With **Read text and labels** checked, the OCR scan then reads and labels what came in. What becomes of each file: - **Images** (PNG, JPEG, HEIC and AVIF, WebP, GIF, BMP, PSD, TGA, TIFF, and camera RAW such as DNG, CR2, CR3, NEF and ARW): an `.mcap` holding the file exactly as it was, a preview turned the right way up (by its EXIF orientation), and where it came from: its file name (searched by `title:`), the camera or phone (`app:`), and the date it was taken, from its EXIF, else a date in its name (`IMG_20240512_101530`, `Screenshot 2024-05-12 at 10.15.30`), else the file's time. It is named `2024-05-12_10-15-30_1a2b3c4d.mcap`, the date and the start of its hash. HEIC, AVIF, WebP and RAW are read by Windows' own codecs: the HEIF, AV1, WebP and Raw Image Extensions from the Microsoft Store. - **Animated GIFs** are images; their thumbnail is marked **GIF**, and the Capture window says how long they run and opens them in the default viewer. - **Videos** (MP4, MOV, M4V, 3GP, MKV, WebM, AVI, WMV, MPEG, MTS and M2TS, FLV), read with FFmpeg (`modules/ffmpeg`, Windows only): an `.mcap` holding the file exactly as it was, of any size (it is copied a few MB at a time, never read whole), and a frame from it a tenth of the way in (at most 3 seconds), turned the right way up, which is what the grid, the Capture window, OCR and labels see. The date is the phone's (Apple's creation date), else one in its name, else the video's own creation time, else the file's time; the phone and its location are kept too. The thumbnail shows a play mark and how long it runs; **Play in the default player** in the Capture window writes the video out to the temporary folder and opens it. - **Snagit `.snagx`:** the same `.mcap` as `snagit-converter mcap` makes, byte for byte, named as the `.snagx` is. - **Duplicates:** an image already in the library, in any of its folders, is not stored again; the Import window lists each with where it is. Importing a folder a second time imports only what is new. - **Skipped:** sidecar files (`.aae`, `.xmp`, `.json`), a GoPro's thumbnails and low-resolution copies (`.thm`, `.lrv`), videos on macOS and Linux, and anything that is not an image or a video, counted by kind. - **Failed:** files that cannot be read or decoded, listed with why. The files imported from are only read: nothing there is changed, moved or deleted. ## Searching Type in the search box above the grid; the grid shows the matches as you type, with "N of M captures" beside the box. Clear the box to see everything again. ### What a search matches - **Every word must match**, in any order: `kicad footprint` finds captures with both. - **Matching ignores case and finds parts of words:** `foot` matches "Footprints". - **A plain word is looked for everywhere:** the OCR text (every model's), the window title, the URL, the app's name, the annotations' text, and the labels. - **"Quoted phrases"** match as one, spaces and all: `"symbol : footprint"`. There is no OR and no NOT: to widen a search, take a word out. ### Prefixes A prefix limits a word, or a quoted phrase, to one field: `title:kicad`, `title:"pcb editor"`. | Prefix | Looks in | | --- | --- | | *(none)* | everything below except `like:`, `folder:` and `has:` | | `ocr:` | the text read from the image | | `title:` | the window title when it was captured; for an imported image, its file name | | `url:` | the web page's address, for browser captures | | `app:` | the app's name and version; for a photo, the camera or phone | | `note:` | the text of the capture's annotations (Snagit text boxes and callouts) | | `label:` | the capture's labels: its best label, and any scoring at least half as much | | `like:` | the image's meaning (below) | | `folder:` | the library folder it is in: `folder:work` | | `has:` | what segmentation found in the image (below): `has:dog` | A word with a colon that is not one of these stays an ordinary word, so `http://` and `C:\` work. ### What is in the image: `has:` `has:dog` finds images that segmentation (the Segmentation window) found a dog in. Only images in folders set to segment, and already segmented, take part. - **The whole class name:** `has:table` is a table, not a dining table; `has:"dining table"` is that. - **The 133 COCO classes:** 80 kinds of thing (person, bicycle, car, dog, cat, chair, laptop, cell phone, book ...) and 53 of stuff (sky, wall, grass, road, floor wood, ...). The Stats window lists those in the images shown. - **Combining:** `has:person has:dog` finds images with both; `has:cat label:"photo"` mixes it with labels. ### Searching by meaning: `like:` `like:"a circuit board under a microscope"` ranks every labelled capture by how close its image is to the phrase, best first: no words need to appear in the image. The label model compares the phrase with the image vector each capture's file holds, so only labelled captures take part. The grid is then in that order instead of by date. - Describe what the image shows: `like:"a graph with several colored lines"` works better than `like:graph`. - Combine it with other words to filter the ranking: `like:"a dark themed code editor" title:visual`. - The phrase is quoted when it has spaces: `like:"a terminal with an error"`. ### Examples | Search | Finds | | --- | --- | | `digikey` | captures mentioning DigiKey anywhere: text, URL, title | | `url:digikey` | DigiKey pages | | `app:kicad` | captures of KiCad, by the app's name (more than `title:kicad`, as a window title need not say KiCad) | | `ocr:"0402"` | images with "0402" in their text | | `title:kicad footprint` | KiCad windows whose text, notes or labels mention a footprint | | `label:schematic` | captures labelled Schematic | | `label:"code review"` | captures labelled Code review | | `note:todo` | captures with an annotation saying TODO | | `like:"a photo of a soldering iron"` | captures that look like that, best first | | `like:"a spreadsheet" app:excel` | Excel captures, the most spreadsheet-like first | | `has:dog` | images with a dog in them (segmented folders) | | `has:person has:"cell phone"` | images with a person and a phone | ## Command line | Option | What it does | | --- | --- | | `-library FOLDER` | Shows the `.mcap` files in FOLDER (and its folders) instead of `D:\McapIndex\mcap`. | | `-import SOURCE DEST` | Imports the folder SOURCE into the library folder DEST, as the Import window does, then exits: 0, or 1 if any file failed. | | `-search QUERY` | Starts with QUERY in the search box. | | `-scan` | Scans the library (the chosen OCR model, and labels when the label model is up) once it has loaded, then exits: 0, or 1 if any capture failed. | | `-stats` | Logs the Stats window's numbers once the catalog has loaded, then exits; with `-search`, for the matches, and the first ten are logged. | | `-segment` | Segments the images not yet segmented in the library folders set to segment (see the Segmentation window), once the library has loaded, then exits: 0, or 1 if any image failed. | | `-erase ocr`, `-erase labels`, `-erase segments`, `-erase all` | Erases that from every file in the library, then exits (for testing). | | `-allow-multiple-instances` | Runs beside another copy instead of bringing that one forward. | | `run_tests` | Runs the test suites (debug builds); with `-headless`, exits with 0 if all passed. | | `-headless` | No message boxes: a failure is logged and the app exits with 1. | | `-version` | Prints the version. | `-import`, `-scan`, `-stats` and `-erase` still open the window, to show their progress; they are meant for scripts and for trying things on a copy of the library: `bin\mcap.exe -library D:\copy -scan`. ## Where things are kept | What | Where | | --- | --- | | OCR readings, labels, image vectors | inside each `.mcap` file (see `modules/Capture_File/README.md`) | | The label list | `%APPDATA%\mcap\labels.config` | | Settings (theme, hotkeys, OCR model) | `%APPDATA%\mcap\settings.config` | | Logs | `logs\\` beside the program | | The hash of every capture's image, for finding duplicates | `.import-index` in the library folder | | The folder each library folder was imported from | `.import-sources` in the library folder | | Which library folders are segmented | `.library-folders` in the library folder | Nothing is written to the folders imported from (`D:\Snagit`, a phone's backup): the library is a converted copy. ## How it works - `src/app/viewer.jai`: the grid and the Capture window. Previews and full images are decoded on worker threads and uploaded a few per frame; textures far from view are released past 1200. Each capture keeps its index for the app's life (Refresh and Delete mark captures removed and append new ones), as the workers, the scan and Stats refer to captures by index. - `src/app/catalog.jai`: what search needs from every file, read from the front of each `.mcap` (its text, OCR, labels and vector, not its images) on the workers at startup; the query parser and matching. - `src/app/background.jai`: searching and the Stats numbers run on a pool of threads, so typing never waits; the grid keeps the last results until the new ones are back, usually within a frame. `like:` phrases are turned into vectors there first. - `src/app/ocr_ui.jai` and `labels_ui.jai`: the engines, made on one background thread (OCR first, then labels); the scan, which decodes each image once and stores its OCR, labels and vector in one rewrite of its file; the erase; the windows. - `modules/OCR`, `modules/Labels`: the models, run with onnxruntime (`modules/onnxruntime`, which also holds the shared session and CUDA code); each has a README. - `src/app/import.jai`: the Import window and `-import`. A thread lists the files, brings the library index up to date (hashing the image in any `.mcap` it has not seen), then imports on several threads; captures written are handed to the grid each frame. - `modules/Image_Decode`: any image file's bytes to pixels, the right way up: stb_image for PNG, JPEG, GIF, BMP, PSD and TGA, Windows Imaging Component for the rest; EXIF (from JPEG, PNG, WebP, HEIF and TIFF); previews; and videos' frames and what their containers say, with FFmpeg (`video.jai`), and how many frames a GIF has. `bin\image-probe FILE...` (`jai first.jai - image_probe`) shows what it makes of files. - `modules/Snagx`: `.snagx` to `.mcap`, copied from SnagitConverter. - `modules/Capture_File`: reading and writing `.mcap` files, with a video's original left on disk. - `modules/Segment`: panoptic segmentation with Mask2Former, run with onnxruntime (see its README and `segmentation-plan.html`). `src/app/segment_ui.jai` is the scan and the Segmentation window; `src/app/library_folders.jai` the per-folder options. - `modules/ffmpeg`: FFmpeg 6.0's bindings and DLLs (`modules/ffmpeg/windows`, in git; copied beside the executable by the build), from the gyan.dev shared build, for reading videos. # The application template mcap is built on a bare Dear ImGui (docking branch) application template, for Windows, macOS and Linux: the window, the frame loop and everything around them, with the app written into `src/app/app_ui.jai`. ## What you get - **Native rendering per platform.** Win32 + Direct3D 11 on Windows; SDL3 + SDL_GPU (Metal on macOS, Vulkan on Linux) elsewhere. On Windows the window comes up on a WARP device in a few milliseconds while the hardware device is created on a thread, then moves across; a lost device is recovered. - **Docking and multi-viewport ImGui**, with a default layout (`assets/data/default_imgui.ini`) and a Reset Layout hotkey. - **One instance.** Launching the app again brings the running one forward, over a shared memory channel (a file lock decides the owner on macOS and Linux). - **Tray / menu bar.** Closing or minimizing hides the window to the tray; the tray icon brings it back or quits. - **Hotkeys** from a `[hotkeys]` section of the settings file, and **global hotkeys** (`[global_hotkeys]`: Windows and macOS) that work while another app is in front. - **A settings file** (`settings.config`) that is written with every key listed, reloaded as soon as it is saved, and merged rather than overwritten when the app saves to it. - **Themes**: built-ins, 36 shipped `.theme` files, and a theme editor that saves new ones. - **Logging** through a multi-producer, single-consumer queue to a per-version log file and an in-app Event Log, plus crash reports. - **Debug panels**: frame times, memory, renderer, fonts (Ctrl+Shift+D, Ctrl+Shift+1). - **DPI awareness**, live resize that keeps drawing, CJK font fallback from the system fonts. - **Background shutdown**: closing hands the remaining work to a thread while the window shows progress, so it never looks hung. - **Asynchronous tests**, compiled into debug builds and run with `run_tests`. ## Building Requires the Jai compiler, beta 0.2.030 or later. ``` jai first.jai # debug build jai first.jai - release # release build jai first.jai - full_release # debug and release jai first.jai - release compile_tests # release build with the test suite ``` Binaries are written to `bin/`, with `assets/` copied beside them. | Platform | Output | Notes | | --- | --- | --- | | Windows | `bin/mcap-debug.exe`, `bin/mcap.exe` | The release build gets the icon and a manifest, and no console. | | macOS | `bin/mcap-macos-debug`, `bin/mcap.app` | Apple silicon (arm64) only, with a native `jai`. The release build is an ad hoc signed `.app`. | | Linux | `bin/mcap-linux-debug`, `bin/mcap-linux` | Needs SDL3 installed (`libSDL3.so`). | `jai check_targets.jai - linux` type-checks the Linux build from any OS without linking it. The macOS check only works on a Mac, since the Objective-C module runs code at compile time. Run any build with `run_tests` to run the test suites (debug builds include them), `-allow-multiple-instances` to skip the single-instance check, or `-version`. For unattended runs (scripts, CI, agents), add `-headless`: the app exits once the tests finish, with exit code 0 if every suite passed and 1 otherwise, and a failed assertion or runtime check is written to stderr, the log and the crash report and exits with 1 instead of opening a message box (`runtime_support_headless` in `modules/Runtime_Support.jai`): ``` bin/mcap-debug.exe run_tests -headless -allow-multiple-instances > test-output.txt ``` `jai first.jai - no_crash_dialog [...]` makes builds for automation: `NO_CRASH_DIALOG` makes them headless whatever their arguments, and everything they build goes to `bin-headless/` instead of `bin/`, so an automated build never writes over the one you run by hand. ## Starting a new app ``` jai fork.jai - [-folder ] [-bundle-id ] [-version ] [-git] [-build] jai fork.jai - ../my-cool-app "My Cool App" -git -build ``` This copies the template (without `bin/`, `.build/`, `.git/` or `fork.jai`) and sets the names below in the copy. The folder name defaults to the app name in lowercase with spaces as `-` (`my-cool-app`), the bundle identifier to `com.example.`, and the version to `0.0.0`. `-git` initializes a repository and commits the copy; `-build` makes a debug build. It also writes `.focus-config`, a [focus](https://focus-editor.dev) project with the build and run commands, which stays out of git since it holds this machine's paths. ## Making it your app 1. Name it at the top of `first.jai`, or let `fork.jai` do it: `APP_NAME` (window title, tray, macOS bundle), `APP_FOLDER_NAME` (executable and data folder; one short lowercase word), `APP_BUNDLE_IDENTIFIER` and `VERSION`. 2. Replace `assets/images/icon.png` (any square size; it becomes the `.exe`, `.app` and window icon) and the tray icons `tray_icon.ico` (Windows) and `tray_icon.png`. 3. Write the app in `src/app/app_ui.jai`: `app_ui_startup` runs once, `app_ui_frame_start` every frame before input, `app_main_window_ui` draws, and `app_ui_shutdown` runs on the shutdown thread. To add a hotkey, declare a `Hotkey` in `src/ui/imgui_keyboard.jai`, add it to `hotkey_actions_build` and `app_check_key_combinations`, bind it in both `assets/data/default_hotkeys*.config` files (the build fails if a default leaves an action unbound), and raise `APP_USER_SETTINGS_VERSION` in `src/App.jai` so existing settings files get the new default. To add a setting, add it to `App_User_Settings`, `Settings_Seen`, `settings_parse` and `app_save_user_settings_file` in `src/ui/imgui_theme_registry.jai`. ## Layout ``` first.jai build: workspaces, build constants, Windows icon/manifest, macOS .app fork.jai copy the template to a new, named project check_targets.jai type-check another OS's build src/App.jai loads everything; global state (app, platform, perf) src/app/ main loop, your UI, background shutdown src/base/ logging, fonts, debug panels, platform layers (base/platform/) src/ui/ ImGui setup, hotkeys, themes and settings, toasts src/graphics/DX11.jai Direct3D 11 device, WARP-first startup, device-loss recovery src/backends/ ImGui's SDL_GPU backend; src/freetype/: ImGui's FreeType loader assets/ fonts, icons, default hotkeys, default layout, themes modules/ Basic (with arenas), Thread, Shared_Memory_Channel, File_Watcher, Secure_Memory, XDG, sdl3, imgui-sdl, imgui-dx11-win32, imgui-source ``` ## Files at run time | File | Location | | --- | --- | | `settings.config`, `imgui.ini`, user themes | `%APPDATA%\` on Windows, `~/Library/Application Support/` on macOS, `~/.local/share/` on Linux | | Logs | `logs//` next to the executable (next to the `.app` on macOS) | | Shipped themes | `data/themes/` next to the executable | ## Platform notes - Global hotkeys are not implemented on Linux (Wayland has no API for them). A desktop shortcut that launches the executable has the same effect, since a second launch brings the running instance forward. - On XFCE the tray icon stays after the window is restored. - `Secure_Memory` (used by `Basic`'s arenas to wipe memory) is built without libsodium; see the note at the top of `modules/Secure_Memory/module.jai` if your app keeps secrets in memory. Uses https://github.com/overlord-systems/jai-imgui as a template.