Avoxkit · Windows / macOS · v0.2.0

Convert, trim, extract —
one window, done.

Avoxkit is a free desktop audio/video toolbox: convert formats, trim clips, extract audio and subtitles, remove image watermarks. Files are processed only on this machine and never uploaded; no sign-in, no membership, no watermark on output.

  • Windows 64-bit · macOS (Apple Silicon)
  • Free · open source
  • Local processing, no upload
Avoxkit trim page: thumbnail strip, waveform and keyframe ticks on the timeline, with a 05:30 to 10:21 range selected

Free

Every feature is free; no membership tiers, and no watermark on output.

Local

Media files are processed on this machine and never uploaded to any server.

No account

No sign-up, no sign-in, no usage data collected.

Features

Six tools, one task queue

Avoxkit is for everyday audio/video work: turning this file into the shape I need. It is not an editor and has no multi-track timeline.

Convert

Change container, codec or resolution. When only the container changes and the target supports the source codecs, the file is remuxed without re-encoding; folder batches skip existing output automatically.

mkv / mp4 / mov / avi / ts / flv / webm / wmv

Trim

Pick start and end points on a timeline with thumbnails, waveform and keyframe ticks, then save that segment. The start snaps to the nearest keyframe, so the time shown is the real start of the output.

Re-encoded output, not lossless cutting

Extract audio

Export M4A or MP3 from video, or copy the original audio track untouched — no re-encoding, fastest path.

M4A / MP3 batch by folder · 128–320 kbps

Extract subtitles

Embedded text tracks export directly to SRT; when there is none, on-device speech recognition generates them.

zh / en / ja / ko / yue · model ~230 MB on demand

Image repair

Detect watermarks automatically or paint the area by hand, then let a model fill the background back in.

LaMa + YOLO26-Seg · model ~290 MB on demand

Tasks and batches

One queue with progress and cancellation; batch groups can be cancelled as a whole and failed items retried. Watch the picture or the audio envelope while a job runs.

Writes to a temp file first, replaces the final file only on success

Convert

Change the container, not the quality

Conversion and remuxing share one page. Every option starts at “same as source”; when only the container changes and the target supports the source codecs, the file is remuxed without re-encoding.

Folder batches are decided file by file, with presets, target size and hardware encoding available when you need them.

Avoxkit convert page: the Sintel film loaded, with container, codec, resolution and target-size options on the right

Repair

Watermarks can be found, and painted away

Automatic mode detects likely watermark regions on-device and fills the background; manual mode paints over the area and repairs only what was painted. Folder batches always run detection.

Repair is generative: the filled background is not guaranteed to match. Keep originals of important images.

Before repair: a red heart watermark on a mountain meadow
Before
After repair: the heart is gone and the grass and scree are filled in by the model
After

Limits

A few edges, stated up front

Trimming is not lossless cutting. Clips are re-encoded as H.264/H.265 plus AAC; the end point is cut by time and can be off by under half a second.

Remuxing is not instant. The file is still read and written in full; cloud drives and HDDs are limited by disk speed.

Recognition and repair models download on demand. About 230 MB for speech and 290 MB for image repair, with SHA-256 checks; nothing downloads until you use them.

No multi-track editing. No timeline projects, tracks or effects; for editing work, use a dedicated editor.

Get

Get Avoxkit

Current release 0.2.0 for Windows 64-bit and macOS (Apple Silicon), under active development.

Build from source: the repository is at github.com/xxxzhou/avoxkit; build the native layer per native/avtool/README.md and package with script/build_windows.ps1 — the output is a portable folder.

Requirements

  • Windows 64-bit; macOS (Apple Silicon)
  • Hardware encoding is detected per machine; falls back to software encoding
  • Speech and repair models download on first use

The UI is built with Flutter; media processing runs on the open-source engine avox; speech recognition uses sherpa-onnx (SenseVoice), image repair uses LaMa and YOLO26-Seg. The project is open source, source on GitHub.

FAQ

Are my files uploaded?

No. All media processing happens on this machine; the app has no telemetry, cloud sync or auto-upload logic. No account is needed and no usage data is collected.

Does it need an internet connection?

Convert, trim and extract all run offline. The speech recognition and image repair models download on first use (about 230 MB and 290 MB); downloads fetch model files only and never read your media.

Which systems are supported?

Currently Windows 64-bit and macOS (Apple Silicon) desktop builds. Conversion covers mkv, mp4, mov, avi, ts, flv, webm, wmv and other common containers.

Windows shows a SmartScreen warning when opening?

The Windows build is not code-signed yet, so SmartScreen may appear on first run: click "More info" and then "Run anyway". This is common for unsigned independent software, and the download can be checked against the SHA-256 sum shipped with the release. The macOS build is Developer ID signed and notarized, so it opens normally.

How is this different from online converters?

Files never leave the machine: no size or count limits, no upload and download queues, and whole folders batch in one go. That matters most for videos with private content.

How is this different from an editor?

Avoxkit is for regular work: convert, trim, extract. There is no multi-track timeline or effects; for editing, use a dedicated editor.

Can it convert MKV to MP4 without re-encoding?

Yes, when only the container changes and the target container supports the source codecs: the file is remuxed without re-encoding. If the target codec differs, it transcodes. Whole folders batch in one go and existing outputs are skipped.

Can it generate subtitles for a video with no subtitle track?

Yes. When there is no embedded text track, on-device speech recognition (Chinese, English, Japanese, Korean, Cantonese) outputs standard SRT; the model (~230 MB) downloads on first use.