What is Concat?
Concat is a desktop and mobile video editor written in Rust that sets out to cover what people open CapCut for: auto-captions, text-to-speech, background removal, keyframes, effects, titles, multi-track cutting and export up to 4K 60 in 10-bit colour, as H.264, HEVC or AV1. A GPU compositor draws the timeline (Metal, DirectX 12 or Vulkan, with a software fallback), and FFmpeg handles media. There is no account, watermark or subscription.
The AI features run locally. Captions use Whisper (Tiny, Base or Small, 78 MB to 488 MB), text-to-speech uses Kokoro, voice cloning uses Pocket TTS or Chatterbox, and cutout has person, object and brush models; each downloads once from Settings and then works offline. For automation, the shipped app runs as an MCP server with `concat --mcp`, and a JSON-RPC API is served on a local socket; a build from source adds gRPC and a command-line tool.
Builds ship for macOS (Intel and Apple silicon), Windows x86_64, Linux x86_64 and ARM, Android, and iOS by sideloading. The project is in beta, under AGPL-3.0, and almost entirely the work of one developer, jub0t.
Is Concat worth it?
Concat aims at CapCut's feature list rather than Premiere's, and it has found an audience quickly. The repository was created on 25 August 2026, has about 12,300 GitHub stars, and shipped v0.2.8 on 20 October 2026, its eighth release in about seven weeks. The v0.2.8 binaries alone show roughly 110,700 downloads on the release page, led by the Windows installer at about 43,000 and the Android APK at about 38,400.
The local-first claim holds up in the documentation. The code signing policy says there is no telemetry and that the app reaches the network only to check for updates or download a model you asked for. The Remote API is off by default and, when on, listens on 127.0.0.1 and answers only callers holding its token. Every optional model together comes to about 3.4 GB, and the system requirements page recommends 16 GB of RAM for 4K timelines and the larger caption models, against a minimum of 4 GB.
The MCP server is the part that sets it apart from Kdenlive or Shotcut. It exposes tools for projects, trimming, splitting, titles, word-level caption styling, keyframes, effects, stabilisation, export and a `preview_frame` that returns the composited frame as an image, so an agent can check its own work. Agents edit their own projects, not the one open in the window; project folders are locked so the two cannot overwrite each other, and agent writes are confined to your home folder unless you widen them.
The risks are maturity and concentration. jub0t has about 1,100 commits; the next contributor has 47. Of 229 open issues and pull requests, the most discussed are export failures on NVIDIA cards with 8 GB of VRAM (#223, #297), slowdowns in v0.2.5 and startup crashes on older or unusual hardware. Windows builds are not code-signed, so Smart App Control refuses them outright, Windows on ARM builds are paused, and the release notes say the Android APK has not been launched by automated tests. For a weekly Reels or Shorts workflow on a decent GPU it is worth a look today; for client deadlines, CapCut's maturity and Kdenlive's history still count for more.
Where it earns its place
A two-person social team cutting daily Reels and TikToks for a consumer brand
Caption and voice short clips without a CapCut Pro subscription or uploading footage
Run Whisper Small (488 MB) for captions and Kokoro for voice-over on each laptop, then export H.264 at 1080p. No licence cost; budget for 16 GB machines if you edit in 4K, and keep a fallback editor while the export-freeze issues on 8 GB GPUs stay open.
A developer building an internal tool that turns recorded product demos into captioned clips
Have Claude Code assemble, caption and export videos from a folder of recordings
Add the server with `claude mcp add concat -- <path to concat> --mcp`; the agent imports media, adds captions, checks frames with preview_frame and calls export_wait. It runs headless with no window, writes only under the home folder by default, and costs only model tokens.
A newsroom video producer handling unreleased interview footage
Edit and subtitle sensitive material on a machine that must not send it to a cloud service
Download the caption and cutout models once, then work offline; the privacy policy in CODE_SIGNING.md says nothing about projects or media is sent. On managed Windows PCs, IT will need to allow an unsigned installer.
Before you commit
- Setup
- Download an installer from the website or GitHub Releases (.dmg, .exe or .msi, .deb, .rpm, AppImage, Arch package, APK). Unsigned builds need a SmartScreen bypass on Windows or an xattr command on macOS. Building from source needs Rust 1.93+, FFmpeg 7+ libraries, cmake and a C++ compiler.
- Integrations
- MCP over stdio (`concat --mcp`, in window or headless mode) for Claude Code, Claude Desktop and other clients; JSON-RPC 2.0 on a local socket from Settings > Automation; gRPC and a CLI in source builds. Imports and exports SubRip subtitles.
- Data and privacy
- No telemetry, per the project's privacy policy; network use is limited to update checks and model downloads. The Remote API is off by default, binds to 127.0.0.1 and requires a token. AGPL-3.0 places no limit on the videos you make, but anyone distributing or hosting a modified version must publish its source.
- Cost at scale
- Free for any number of seats. The real cost is hardware: 16 GB of RAM and a Metal, DirectX 12 or Vulkan GPU recommended for 4K, and about 3.4 GB of disk for every optional model.
- Limits
- 229 open issues and pull requests, with recurring export failures on 8 GB NVIDIA cards and startup crashes on some systems. Windows ARM builds are paused; macOS 12 is the minimum, and some AI features on Intel Macs need 13.4. Agents cannot yet edit the project open in the window (planned in #337).
What's good
- Captions, TTS and cutout run offline after one download
- No account, watermark, upload or telemetry
- Built-in MCP server with a frame preview for agents
- One project format across desktop and Android
What's not
- Beta with open crash and export-freeze reports
- One developer writes nearly all the code
- Unsigned Windows builds; Smart App Control blocks them
- Android and iOS builds are lightly tested
Best for
Short-form creators and developers who want CapCut-style editing and agent automation with nothing leaving the machine.
Concat alternatives
Concat: common questions
- Is Concat worth it?
- Concat already does most of what people use CapCut for without an account or an upload, but it is a two-month-old beta with one main developer and the crash reports to match. We rate it 3.8 out of 5.
- How much does Concat cost?
- Free and open source (AGPL-3.0). No paid tier, account or watermark; the project asks for donations on Patreon.
- What are the downsides of Concat?
- Beta with open crash and export-freeze reports; one developer writes nearly all the code; unsigned Windows builds; Smart App Control blocks them; android and iOS builds are lightly tested.
- Who is Concat best for?
- Short-form creators and developers who want CapCut-style editing and agent automation with nothing leaving the machine.
- How much does Concat cost for a team or in production?
- Free for any number of seats. The real cost is hardware: 16 GB of RAM and a Metal, DirectX 12 or Vulkan GPU recommended for 4K, and about 3.4 GB of disk for every optional model.
- Is Concat safe for company data?
- No telemetry, per the project's privacy policy; network use is limited to update checks and model downloads. The Remote API is off by default, binds to 127.0.0.1 and requires a token. AGPL-3.0 places no limit on the videos you make, but anyone distributing or hosting a modified version must publish its source.
- What are the best Concat alternatives?
- The closest alternatives we have reviewed are CLI-Anything, VideoAgent and ComfyUI.
Answers checked Oct 28, 2026.