Kando（看懂）

English · 简体中文 · 日本語 · 繁體中文 · 한국어

Kando is a desktop app for learning a language from videos in that language. Paste a YouTube or Bilibili link, or drop a video file on the window: Kando downloads the video with yt-dlp, transcribes the speech on your own computer with whisper.cpp, translates each line through an OpenAI-compatible model, and plays the original video with the original line and its translation on screen together. It is written in Go and draws its native interface with MyGo and Breeze UI; there is no web view.

The interface is in Simplified Chinese for now. The labels quoted below, such as Settings › 语音识别, are the app's own.

Features

• Add a video by pasting a link from YouTube, Bilibili or any other site yt-dlp supports, or pick a local file (mp4, mov, mkv, webm and other video files) or drag it onto the window. Local files are transcribed where they are, not copied.
• Jobs run one after another through download, transcription, translation and export. A failed job says why, links to the setting that fixes it, and can be retried or removed.
• Speech is recognized on your computer with whisper.cpp, which detects the spoken language. Settings › 语音识别 lists seven whisper models, from Tiny (77.7 MB) to Large v3 (3.1 GB), and downloads the one you choose from Hugging Face.
• Lines are translated by EveryAPI by default, or by any OpenAI-compatible chat API: Settings › AI 服务 has presets for OpenAI, OpenRouter, Groq, DeepSeek, xAI, Mistral, Qwen (DashScope), Kimi, Zhipu GLM, SiliconFlow, Ollama and LM Studio, plus a custom endpoint. The target language, a context-aware mode that lets the model read the preceding lines, and the prompt can be changed in Settings › 翻译.
• Each finished video gets two SubRip files next to it: <video>.srt with the original lines and <video>.bilingual.srt with each translation above its original line (video.srt and video.bilingual.srt for a downloaded video).
• The player shows the subtitles as bilingual, original only, translation only or off, with a delay you can adjust in 0.1-second steps and a style you can change. Its side panel lists every line with search (click a line to jump to it), splits the video into 4 to 12 clips of about 30 seconds, and shows the film's files and source. Each film resumes where you stopped watching.
• Libraries are .kando folders of ordinary files you can manage in Finder or a file manager. You can keep several and switch between them from the sidebar, which also filters a library by source: YouTube, Bilibili, other sites and local files.
• Optional upload of a film's video and bilingual subtitles to your own object storage (Cloudflare R2, Amazon S3, Alibaba Cloud OSS, Tencent Cloud COS or any S3-compatible store), from the player, after signing in with Google. See Configuration.
• Keyboard shortcuts for playback, subtitles and windows, listed in a cheat sheet (⌘/) and in Settings › 快捷键.

Install

Download

Builds for macOS, Windows and Linux will be attached to GitHub Releases once a version is published. They are not signed with a Developer ID or code-signing certificate, so the system warns the first time you open Kando:

• macOS: on macOS 15 and later, open Kando once, then choose System Settings › Privacy & Security › Open Anyway. On macOS 14 and earlier, Control-click Kando in Applications and choose Open. Or remove the quarantine flag in Terminal: xattr -dr com.apple.quarantine /Applications/Kando.app.
• Windows: in the SmartScreen dialog, choose More info › Run anyway.

Build from source

With Go 1.27.3 or later:

git clone https://github.com/yldm-tech/Kando.git
cd Kando
make build

make build runs go tool mygo build apps/kando for the platform you are on and writes the result under apps/kando/build/<os>-<arch>/: on macOS an ad-hoc signed Kando.app and a .dmg, on other platforms an executable (Linux also gets a .deb, a .tar.gz with an install script and a PKGBUILD; Windows an installer made with NSIS, which mygo downloads when it is not installed). Without make, as on a stock Windows, run go tool mygo build apps/kando from the repository root.

Requirements

Kando runs these programs; it does not install or update them.

Program	Used for
yt-dlp	Downloading online videos. Kando asks for it for every job, local files included.
ffmpeg	Extracting the audio, grabbing covers and decoding the picture in the player. On macOS the sound plays through ffplay from the same install, or afplay.
whisper-cli (whisper.cpp)	Speech recognition.
A whisper.cpp ggml model	Download one in Settings › 语音识别, or point KANDO_WHISPER_MODEL at a model file.

On macOS with Homebrew:

brew install yt-dlp ffmpeg whisper-cpp

On Windows and Linux, install the same programs with your package manager and make sure they are on PATH. Besides PATH, Kando looks in /opt/homebrew/bin, /usr/local/bin and /home/linuxbrew/.linuxbrew/bin, because apps opened from Finder or a desktop launcher get a minimal PATH. Settings › 通用 shows which programs were found, and 重新检测 looks again.

On Windows and Linux the player currently shows the picture without sound; audio playback is implemented on macOS only.

Getting started

1. Open Kando. With no library yet, it offers 新建资料库… (a new library, named 影视库 in ~/Movies on macOS or ~/Videos elsewhere unless you change it) and 打开已有资料库… (an existing one).
2. Open Settings (⌘,). In 通用, check that ffmpeg, yt-dlp and whisper-cli show as 已安装.
3. In 语音识别, click 下载 next to a whisper model; Kando selects it when the download finishes. The one marked 推荐 is Large v3 Turbo（压缩版）, 574 MB. Models that Vimpo or hyperframes already downloaded can be selected without copying.
4. In AI 服务, set up translation:
   • EveryAPI (the default): sign in once with the EveryAPI CLI, everyapi auth login. Kando asks the CLI for the key (everyapi auth credential), so it uses the same login, gateway region and key refresh as the CLI, and you do not paste a token. The default model is claude-haiku-5-5; everyapi models list shows the others. You can also paste an EveryAPI key into API Key instead.
   • Another provider: choose it under 服务, or choose OpenAI 兼容 and enter its Base URL, then fill in 模型 and API Key (local servers such as Ollama and LM Studio need no key). 测试连接 sends a short request to check the settings.
5. In 翻译, pick the default target language: 简体中文 (the default), 繁体中文, English, 日本語, 한국어, Español, Français or Deutsch.
6. Back in the window, paste a link into the top bar and click 开始处理, or click 本地视频 or drag a video onto the window. The status line at the bottom of the sidebar says 已就绪，可以开始处理 when everything a job needs is in place, and otherwise names what is missing.

When the translation provider answers with a rate limit or a server error, Kando sends the request again, up to three attempts in all.

Configuration

Everything below except the Google client id and secret can also be set in the app (the library from the sidebar, the rest in Settings). An environment variable, when set, takes the place of the value saved in the app.

Variable	Default	Meaning
KANDO_WHISPER_MODEL	the model chosen in Settings › 语音识别	Path of a whisper.cpp ggml model file, such as ggml-small.bin.
KANDO_LLM_BASE_URL	EveryAPI, https://api.everyapi.ai/v1	Base URL of an OpenAI-compatible provider, such as https://api.openai.com/v1, or http://localhost:11434/v1 for Ollama. When it is set, the signed-in EveryAPI account is not used.
KANDO_LLM_MODEL	claude-haiku-5-5 on EveryAPI	Translation model. With a custom provider there is no default: set it here or in Settings › AI 服务.
KANDO_LLM_API_KEY	none	Bearer token for the translation provider. It overrides the signed-in EveryAPI account.
EVERYAPI_API_KEY	none	EveryAPI key, used when KANDO_LLM_BASE_URL is unset and KANDO_LLM_API_KEY is empty, instead of the CLI login.
KANDO_OUTPUT_DIR	the library opened last	Library folder to open at launch, where new videos and subtitles go.
KANDO_GOOGLE_CLIENT_ID	none	Google OAuth client id for signing in on Settings › 账号与上传. Without it the sign-in button is disabled. It is never saved to the settings file.
KANDO_GOOGLE_CLIENT_SECRET	none	Optional client secret for the same OAuth client.
KANDO_STORAGE	r2	Where uploads go: r2, s3 (or aws), oss (or aliyun), cos (or tencent), compat (or s3compat). Any other value means R2.
KANDO_R2_ENDPOINT	none	Cloudflare R2: an account id or a full https:// endpoint.
KANDO_R2_BUCKET, KANDO_R2_ACCESS_KEY, KANDO_R2_SECRET	none	Cloudflare R2 bucket and keys.
KANDO_S3_REGION	none	Amazon S3 region name, such as ap-northeast-1.
KANDO_S3_BUCKET, KANDO_S3_ACCESS_KEY, KANDO_S3_SECRET	none	Amazon S3 bucket and keys.
KANDO_OSS_REGION	none	Alibaba Cloud OSS region name, such as cn-hangzhou (cn-hangzhou-internal for the internal network).
KANDO_OSS_BUCKET, KANDO_OSS_ACCESS_KEY, KANDO_OSS_SECRET	none	Alibaba Cloud OSS bucket and keys.
KANDO_COS_REGION	none	Tencent Cloud COS region name, such as ap-guangzhou.
KANDO_COS_BUCKET, KANDO_COS_ACCESS_KEY, KANDO_COS_SECRET	none	Tencent Cloud COS bucket (its name includes the APPID, such as video-1250000000), SecretId and SecretKey.
KANDO_S3COMPAT_ENDPOINT	none	Full https:// endpoint of another S3-compatible store.
KANDO_S3COMPAT_REGION	none	Signing region of that store; MinIO usually uses us-east-1.
KANDO_S3COMPAT_BUCKET, KANDO_S3COMPAT_ACCESS_KEY, KANDO_S3COMPAT_SECRET	none	Bucket and keys of that store.
KANDO_S3COMPAT_PATH_STYLE	off	1, true, yes or on turns on path-style addressing, which MinIO needs; 0, false, no or off turns it off.

Uploading is manual: open a film and click 上传到 R2 (or the store you chose) in the player. Kando uploads the video and its bilingual subtitles under <Google account email>/<library>/<film folder>/<file>; the sign-in only says who uploaded. Videos are never uploaded automatically.

Libraries and settings

A library is a folder whose name ends in .kando. Each downloaded film gets its own folder inside it, <title> [<video id>]/, holding video.mp4, cover.jpg, video.srt and video.bilingual.srt. They are ordinary files: Kando watches the library folder, so videos added or removed in Finder or a file manager show up in the library. On macOS a .kando folder gets a folder icon with the Kando mark in Finder.

Removing a film from the library moves the film's folder to the Trash, or only its video and subtitles when that folder holds other videos or folders too; a local video outside the library folder stays where it is and only leaves the list. Removing a library from the list leaves its folder on disk.

Kando keeps its own files in the user config directory, ~/Library/Application Support/Kando on macOS, %AppData%\Kando on Windows and ~/.config/Kando on Linux:

• library.json: the jobs, finished and in progress, and where each video stopped playing.
• settings.json: the settings, including the API key, the object storage keys and the Google refresh token, in plain text. The file is readable only by your user.
• window-state.json: window sizes and positions.
• models/: whisper models downloaded in Settings › 语音识别.

Cover images of videos that are not named video.* are cached in ~/Library/Caches/Kando/posters on macOS, %LocalAppData%\Kando\posters on Windows and ~/.cache/Kando/posters on Linux.

Upgrading from LingoCut

Kando was called LingoCut, and nothing reads the old names. To keep your libraries and settings:

1. Quit the app.
2. Rename the config folder from LingoCut to Kando: ~/Library/Application Support/LingoCut on macOS, %AppData%\LingoCut on Windows, ~/.config/LingoCut on Linux. It holds library.json, settings.json, window-state.json and the downloaded models.
3. Rename each library from X.lingocut to X.kando, after renaming the config folder and before launching. On macOS and Linux the saved inodes let the app follow every renamed library. On Windows the app can only match one missing library to one new .kando folder in the same parent folder, so rename one library per folder at a time and launch the app once after each rename.
4. Rename each environment variable from LINGOCUT_* to KANDO_*.
5. Launch Kando and choose the speech model again in Settings › 语音识别: the model file moved with the folder and shows as downloaded, but the saved choice still points at the old path.
6. Delete what is left of the old name: lingocut.app and lingocut Dev.app; the poster cache ~/Library/Caches/LingoCut (%LocalAppData%\LingoCut on Windows, ~/.cache/LingoCut on Linux); on Linux the lowercase ~/.config/lingocut; and, if you ran make dev, ~/Library/Application Support/lingocut Dev.

Development

Requires Go 1.27.3 or later.

make test   # go test for packages/core, packages/media and apps/kando
make dev    # run the app with live reload
make build  # build the app for this platform into apps/kando/build

make dev runs MyGo's mygo dev patched by tools/mygo-dev: on macOS, a reload keeps the running window until the new build shows its own, instead of quitting it first, so saving a change does not look like the app quitting. If the patch stops applying after a MyGo upgrade, it warns and falls back to the stock go tool mygo dev; update tools/mygo-dev/dev.go.patch against the new cmd/mygo/dev.go then. The script behind it, tools/mygo-dev/dev.sh, is a POSIX shell script.

The screenshot tests render the main scenes to PNG files from a fixed demo library, in an empty fake home folder with fake ffmpeg, yt-dlp and whisper-cli, so they do not read your settings, libraries or tools:

KANDO_SCREENSHOTS=/tmp/kando-shots go test ./apps/kando/... -run TestScreenshots

Repository layout:

• apps/kando: the app. Each screen has its own directory (sidebar/, films/, film/, settings/, account/, cheatsheet/); view.go connects the window to them. shelf/ is the job record, cloud/ is Google sign-in and object storage upload, and brand/ is the Finder icon of a library.
• packages/core: platform-independent jobs, stages and link parsing.
• packages/media: the media pipeline: download, transcription, translation, subtitle export, covers and libraries.
• tools/mygo-dev: the patched mygo dev behind make dev.

AGENTS.md describes how the interface is split into components and the rules each area follows; read it before changing the UI.

CI and releases

The CI workflow builds and tests pull requests and main. The Release workflow runs after CI passes on main, or by hand: it picks the next version from the Conventional Commits since the last tag, commits it to main as chore(release), tags it, builds packages for macOS, Windows and Linux without a Developer ID or code-signing certificate, and publishes a GitHub release with their checksums.

License

MIT