kiri: a Tauri capture app that keeps the un-redacted original, and ships unsigned
A fast, local-first workspace for screenshots, annotation, OCR, and screen recording.
At a glance
- What is it?
- A local-first screenshot, annotation, OCR and recording tool for three desktop platforms, written in Rust with a web frontend. The page is candid about platform limits, and two of them are worth reading before you install: Linux loses audio and video editing entirely, and neither desktop build is signed.
- Who is it for?
- kiri is a well-built capture tool with an unusually honest account of its platform limits, and both of the things that should give you pause are on the page rather than hidden. Linux is a reduced target: one display under Wayland, silent recordings, no click highlight and no video editing.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 10 days ago.
- What is it written in?
- Mainly Rust, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 21, 2026, and from our analysis. They are not legal advice.
Star history
- yuxino/kiri575 ★ · +143
| Repository | 2026-09-17 | 2026-11-21 |
|---|---|---|
| yuxino/kiri | 432 | 575 |
143 Stars in the last 65 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
Annotated screenshots keep the un-redacted original on disk
The privacy section is three sentences long, and the third one is the one that matters most.
Captures and media processing stay local. Remote OCR is optional and asks before each upload. And editable screenshots retain an original image locally, including pixels covered by annotations.
That last clause is the whole privacy model in one sentence. If you draw a box over a name in a screenshot, the box is an overlay, and the pixels underneath are still in the original file on your disk. The same applies to the mosaic tool.
From an engineering standpoint this is the correct choice. Destructive redaction is irreversible, and a user who later decides the crop was too aggressive cannot get the pixel back. The cost is that redaction is not redaction in the sense a compliance officer means by the word.
So the practical rule is that a mosaicked screenshot should be treated as a screenshot with something drawn on it, not as a sanitised image. If the underlying image cannot be in the library at all, the answer is to crop rather than to cover, and to delete the capture rather than to annotate it.
The page links a privacy policy document, and a Chinese-language version of it sits in the repository root alongside the English one.
Kiri discards the screenshot it takes to ask for permission
There is one privacy behaviour on the permission path that is genuinely unusual, and it is stated almost in passing.
On GNOME Wayland, if the first capture shows no permission dialog, you are told to open the library, choose Request Access in the error banner, allow screenshot access in GNOME's own dialog, and retry the capture. Then the sentence that makes this worth noting: Kiri discards the authorization image.
That screenshot exists only because the capture API required a frame before it would show a permission prompt. It necessarily contains whatever was on screen at that moment. A tool that keeps it would be silently accumulating a second copy of the user's entire desktop, taken at moments they did not initiate.
So the authorization capture is destroyed rather than routed to the library, while a capture the user actually takes is saved. That distinction between a frame the app had to take and a frame the user asked for is the kind of thing most tools do not make explicitly, and making it explicitly is what makes the rest of the local-first claim checkable.
The same section of the page covers a different case: on macOS, if the display layout, resolution or scale changes after a region was selected, you are told to start a new capture before recording, and to stop and save first if a recording is paused. That is a race between the compositor and a selected region, and it is handled by telling the user rather than by re-measuring.
Neither desktop build is signed, and the bypass is documented
The download table lists three platforms and then, immediately below it, two signing disclosures.
On macOS the app is not notarized, and the instruction for a blocked launch is to open System Settings, go to Privacy and Security, and choose Open Anyway. On Windows the packages are not Authenticode-signed, so SmartScreen may warn.
Both of those are the operating systems asking the user to take responsibility for running an unverified binary, and both instructions are about how to say yes rather than how to verify. There is no checksum to compare and no signature to inspect on the page.
That is worth weighing against what the application does. A screen capture tool holds keystrokes' worth of context, whatever is in the clipboard at the moment you press the shortcut, and by default every capture lands in a library that the same tools can search, tag and export.
The macOS permission story is separate from signing and is properly gated: Screen and System Audio Recording, plus Input Monitoring for click highlights, plus macOS 15 or newer for microphone recording. So the OS-level permissions are real. Signing is a different control, and it is the one that is absent.
The release page and a Linux package are the offered downloads, and the page also notes that the Windows portable build keeps its settings and captures in the user profile.
Linux gets one display, silent recordings and no video editing
The Linux paragraph is short and lists everything that is missing.
Wayland capture supports one connected display. Recording is silent, with no click highlights and no video editing.
Add to that the video editing feature itself, which the capability list scopes to macOS and Windows only, and the picture is a platform with capture, annotation, text recognition and QR reading but a substantially reduced media pipeline.
Silent recording is the sharpest difference. On the other two desktop platforms you can record system audio, the microphone and a visible click highlight together. On Linux you get the picture with no sound at all.
The build target is narrow too. The table offers Ubuntu 24.04 on x64 with GNOME or X11 and nothing else, so there is no Debian, Fedora or Arch package listed. And there is no auto-update on Linux: the About settings path is for the macOS and Windows installer builds, and Linux users download a new package from the releases page.
So the update story and the feature story agree with each other. Linux is the supported platform where the media features stop.
Even on the other two, the update path splits by packaging rather than by platform, since the Windows portable build is in the same download-a-new-package group as Linux.
Escape is context-sensitive, and the first press leaves the text edit
The keyboard section is where the page gets precise about a class of small bug that is easy to get wrong.
While typing an annotation, the undo shortcut reverses text and a modified enter adds a line. Escape leaves the text edit first, and a second Escape cancels the capture.
So the cancel gesture is two presses, not one, whenever a text annotation is open. A user who reaches for Escape to abandon a capture, having just clicked into a text box, will instead leave the text box and find themselves still in capture mode.
The same layered-escape idea appears elsewhere with the opposite polarity, where closing an edited saved image offers three choices rather than one: save, discard, or keep editing when there are unsaved changes. No single keystroke is given a destructive default.
Capture confirmation is otherwise one key each. Enter confirms a screenshot, Escape cancels, and the three choices at the top of the flow are Screenshot, Record and OCR, with QR recognition offered from the screenshot toolbar.
The launch shortcut also has three forms rather than one. A modifier combination with the command key on macOS, the control key on Windows and Linux under X11, and on Wayland either the Capture button or the same shortcut bound through desktop settings using a command-line capture flag.
Three READMEs, two privacy policies, and the demo is delimited by version markers
The repository root carries more documentation than a project of this size strictly needs, and the duplication is visible rather than hidden.
There are three readme files. The main one is the English version, since the header links a Chinese file for that language and bolds the word English for itself. So a second file named for English sits next to a file that is already English.
There are two privacy policy documents, an English one and a Chinese one. Only the English one is linked from the readme, and the Chinese one is reachable only by browsing the repository.
The demo video, by contrast, is wrapped in a pair of HTML comments carrying a version marker and its closing counterpart. That is a strong signal the video block is regenerated by a script and swapped in rather than hand-edited, which is the right way to keep a media embed current.
The rest of the documentation is a normal set of separate documents: a contributing guide, a security policy, a roadmap, a video editing guide, a documentation index, a Linux guide referenced twice from the download and capture sections, and a machine-readable agent instruction file.
The single credit in the thanks section is for the initial Linux support, attributed to a named contributor and a specific pull request, which is worth noting given how much of the platform story above depends on it.
The release gate is one node --test command naming twenty-nine files
The scripting in the project manifest is worth reading, because it describes how the project checks itself.
The release tools test is a single command. It runs the built-in test runner over twenty-nine test files plus one installer theme script, and the list is written out longhand in the manifest rather than globbed.
There is no test framework in the dependency list. This is the runtime's own runner, invoked directly, on plain module files. Which means the twenty-nine names are a readable map of what the project considers worth checking, and they are worth reading as one.
Roughly half are feature areas: the editor and its document, crop, image edit and QR paths, text composition and layout, annotation interaction locking and the annotation project, the video trim, effects and save paths, the library card, the viewer, OCR selection commit and OCR history, QR results, the toolbar and countdown clock.
The other half is process: capture runtime boundary, repository hygiene, release version, the updater manifest and flow, dock settings, compact translations, resource URLs, notice icons, and the test runner's own wrapper.
A separate one-command script covers development signing and re-runs one of those files. And the macOS release packaging is a shell script while every other script is a JavaScript module.
Icon verification gates both the dev server and the build
Two manifest details are worth calling out because they change what a contributor experiences.
The first is that icon verification runs before the development server and before the build. Both commands begin by running it, then the build additionally typechecks. So a wrong or missing application icon stops you before you see a page, not at some later packaging step.
That is a stricter choice than most projects make, and it makes sense here because the icon set is generated rather than drawn by hand. There is a generate command alongside the verify one, which means the intended workflow is that the icons are produced from a source of truth and the check exists to catch a build that skipped that step.
The second is the split between two update mechanisms. The project ships an updater plugin and an in-app check, and the readme says the installer builds on macOS and Windows use it from the About settings. Linux and the Windows portable build do not, and download from the releases page instead.
So the same application has two update paths on the same platform depending on whether it was installed or unpacked, and one path on Linux. A portable build that never checks for updates will keep running whatever version it was unpacked with.
Editorial conclusion
kiri is a well-built capture tool with an unusually honest account of its platform limits, and both of the things that should give you pause are on the page rather than hidden. Linux is a reduced target: one display under Wayland, silent recordings, no click highlight and no video editing. And the desktop builds are unsigned, so macOS and Windows each ask you to override their gate before the app is allowed to see your screen. Read the privacy section carefully as well, because annotated screenshots deliberately keep the un-redacted original on disk, which is the right engineering decision and the thing to remember the next time you mosaic something.
Frequently asked questions
Which platforms does kiri support?
macOS 14 and newer on Apple silicon and Intel, Windows 11 on x64, and Ubuntu 24.04 on x64 with GNOME or X11. The downloads offered are a universal disk image, a Windows installer or portable zip, and a Debian package installed with apt.
Is the kiri app signed or notarized?
No. It is not Apple-notarized, so a blocked launch requires opening it from the privacy and security settings, and the Windows packages are not Authenticode-signed, so SmartScreen may warn. The page gives the bypass steps rather than a verification method.
What does the Linux build of kiri leave out?
Wayland capture supports one connected display, recording is silent with no click highlights, and video editing is described as macOS and Windows only. Linux also downloads updates from the releases page rather than checking in the app, and only an Ubuntu 24.04 package is offered.
Does kiri send screenshots anywhere?
Captures and media processing stay local, and remote OCR is optional and asks before each upload. Editable screenshots retain an original image locally, including the pixels that annotations cover, so a mosaic is an overlay rather than a redaction.
How are kiri updates delivered?
The macOS and Windows installer builds check for updates from the about settings. Linux and the Windows portable build download a new package from the releases page instead, and portable settings and captures stay in the Windows user profile.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/yuxino-kiri)Star history chart for your README
[](https://hysenlabs.com/projects/yuxino-kiri)Was this useful?