ComeCut is a browser-first video editor with an AI dubbing pipeline

An editor that opens in a browser with no account

ComeCut presents itself as a free, lightweight, AI-powered video editing suite for web and desktop, with features inspired by editors such as CapCut, and the project description spells out the promise plainly: free for the web, the desktop and more. The web version is the primary artefact, since the repository root is an `index.html` alongside `en/`, `web/`, `assets/` and `update/` directories, and the online demo is published as a canary build on GitHub Pages. The commercial angle is that nothing asks you to register. The documentation states there is no registration, that you start creating immediately with no account required, and that there are no hidden costs and unlimited usage. For a person who wants to cut a clip, add a subtitle and export, that removes every step between installing software and editing. It also means the editor is a web application in the strict sense, with the browser doing the work and your project data living in local storage, which is a design decision with consequences for large projects and for anyone who wants the source of their video to stay on their own machine.

The one-click dubbing chain is the actual product

The headline feature is fully automatic video translation, and the roadmap describes it as a pipeline: speech recognition through ASR, then subtitle translation, then speech synthesis through TTS, then video synthesis. The two items checked off in the roadmap are automatic subtitle generation from audio tracks and one-click localized dubbing, for example English to Mandarin. What the 2026-10-05 milestone added is the ability to turn any video into a Chinese or English voiceover with a single click, and the preview version supports API integration through `audio.cpp` using OpenAI's standard protocols. The model list is named: index-tts2 and index-tts2.5, Breeze-TTS2, OmniVoice and Qwen3-ASR. The reason that list matters is that the model choice determines the voice quality and the language coverage of the dubbing, and the editor is deliberately not shipping those models itself. The separate audio and restore functions on any clip, reachable from the right-click menu, are the more conventional editing features in the same release.

Models are added by API, not shipped, which is why the package shrank

This is the design decision that explains the rest of the project. The 2026-10-05 notes describe an attempt to integrate large models directly, which caused the package size to balloon because of the dependencies involved, and then a change of approach: the models were decoupled and local calls go through APIs instead. The wording in the notes is unapologetic about it. When you add a model, it appears in the right-click menu automatically, so a model such as Qwen3-ASR or IndexTTS2.5 becomes a per-clip action once it is registered. The acknowledgement section credits `audio.cpp` by ShugoAI as the local audio inference engine, and the underlying project is linked. For the AI assistant, the supported protocols are named too: Gemini and OpenAI, including Qwen 3.8, as well as local Ollama and LM Studio. The practical consequence is that the editor itself stays small and installs quickly, and the cost moves to you, because you either point it at a hosted API and pay per call, or you run the models yourself and manage the hardware.

An agent chat that edits the timeline

The second AI surface is a conversational assistant inside the editor. The documentation lists what it can do: locate footage, edit, adjust volume, modify subtitles, translate subtitles, and generate images and videos using ComfyUI, all from the agent chat. The example list of commands is more revealing than the marketing, because it is written as editor operations rather than questions. It can translate the current subtitle into English, find and correct typos in subtitles, find music and add it to the track, create an image through ComfyUI, find a specific subtitle and position it on the timeline, delete or trim a clip, add music and adjust the volume to a percentage, and save the project. The last entry in that list is telling: you can delegate anything to it. This is an editor whose intent parsing is allowed to touch project state, which is a large design surface, and the fact that saving the project is a delegated action means a wrong interpretation can be persisted rather than previewed.

Subtitles, cutout, GIF export and the other editing basics

The non-AI feature set is what determines whether the editor is usable at all. Subtitle handling is the strongest area, with smart translation for SRT, VTT and LRC formats, support for multiple languages in one click, and dual-language display. The cutout feature is described as integrated and one-click, with hair-level cutout claimed, and GIF export is built in for one-click generation of GIFs and emoji animations. The desktop version for Windows has been compiled and is available for download, and the platform table in the README marks Windows, macOS and Linux all as beta. The documentation also notes a privacy position: no data collection, no tracking, no personal information collected, and project data stored in the browser's local storage, with the sentence about local storage cut off at the end of the README. For an editor whose selling point includes local processing, that storage model is the piece to understand, because browser storage is per-origin, per-browser and per-machine, and clearing site data takes the project with it.

A 30 day expiry and a beta label on every platform

Two constraints deserve more attention than the feature list gives them. The first is stated in a block quote near the top: to keep you up to date, the developer version is set to expire automatically after 30 days. This is the canary build, and it means the online demo is not something you can rely on as a permanent installation without expecting to re-authenticate or re-fetch periodically. It also tells you something about how the project manages binaries, since the build has to check in with the project to stay valid. The second constraint is the beta status. The version badge says beta, all three desktop platforms are marked beta, and the community section asks contributors to wait for the project to stabilise before making major contributions. The release cadence is fast rather than steady: v0.1.1 on 2026-06-05 through v0.1.9 on 2026-10-07, then v0.2.0 and v0.2.1 in the following two weeks, with the last push on 2026-10-20. A project that ships two releases in two days and labels every platform beta is one to try before you commit a production pipeline to it.

AGPL terms and an open roadmap

The licence is AGPL 3.0, which is the part of the project most likely to affect a decision about adoption, so it deserves a plain statement rather than a recommendation. AGPL extends copyleft to network use, so if you modify the editor and then offer it to users over a network, the modified source has to be offered to those users as well. Running the unmodified project yourself, or self-hosting the public build, is a different matter from distributing a fork. The project also publishes a bilingual README, with an English version and a Simplified Chinese one, and a `guide/` directory alongside `example/`, `hashmap.json` and `vp-icons.css`, so the interface strings and guides exist in both languages rather than one being a translation afterthought. The roadmap shows what is not finished yet: an AI creative studio integrating Seedance-2.0, Veo3.1 and Sora2 for AI short dramas and comics is still unchecked. The last push to the repository was on 2026-10-20, and the v0.2.1 release came the day before.

Editorial conclusion

Adopt ComeCut if you want to edit video in a browser without an account, and you are willing to point the AI features at your own model endpoints, including local ones such as Ollama or LM Studio. Do not adopt it if you need a stable long-term desktop build, since the Windows, macOS and Linux packages are all marked beta and the developer version expires after 30 days, or if you cannot work under the AGPL 3.0 licence. Before you commit, check whether the model you plan to use is reachable through the API settings, and keep the 30 day expiry in mind if you plan to host the developer build yourself.

Frequently asked questions

Official sources

1. juntaosun/ComeCut on GitHub
2. License: AGPL-3.0
3. Project website
4. README
5. Releases