Concat is the free CapCut that agents can drive. The README promises an MCP server; the docs say it doesn't exist yet

The clip that sent me here is thirty-six seconds long. An account called github.signals posted it on TikTok on October 8 under the name WolfCut: a free, open-source CapCut alternative with no watermark, no account, no subscription, a Rust engine, and its own bundled FFmpeg. It has 466,000 views and 31,000 likes as I write this. The repo it points to, jub0t/wolfcut, now redirects to jub0t/concat, because the project renamed itself. If you go looking for WolfCut you will find the video and not the software.

I render most of my own video from HTML with HyperFrames, and the gap in that pipeline is not motion graphics. It is the boring cut: trim a talking-head clip, burn captions on it, drop a title, export vertical. CapCut does that and charges for it in pieces. So an editor that does it for free, offline, and exposes an API an agent can call is exactly the thing I would want. That is why I read the whole repo before installing anything, and why this is a reading of the project rather than a review of it. I have not shipped a video with Concat yet. I will say what I checked and what I did not.

No affiliate link, no sponsor: Concat is AGPL-licensed software with no paid tier to sell you, and I have no relationship with the project. Every number below comes from the GitHub API, the release page or the project's own documentation, checked on October 15, 2026, and all of it is linked at the bottom.

What it is, in the project's own words

Concat describes itself as a free, open-source video editor and CapCut alternative for macOS, Windows, Linux and Android that covers "what people actually open CapCut for": auto-captions, text-to-speech, background removal, keyframe animation, effects and titles, multi-track cutting, 4K export. The catches it lists are the ones that matter to people who make short video for a living: no watermark, no account, no subscription, no upload. Everything runs on a native Rust engine with a GPU compositor, and the AI models for captions, voices and cutout download once from Settings and then work offline.

The numbers behind the hype are modest by TikTok standards and healthy by open-source ones. The repository was created on August 25, 2026. This morning the GitHub API reports 5,589 stars and 550 forks. Version 0.2.7 shipped on October 13, a week after 0.2.6, and the project calls itself beta in the README: "it works, and it still has edges." The release assets tell you who is actually downloading it. For 0.2.7 the Windows installer had 5,950 downloads, the Android APK 5,567, the iOS build 1,213, the Apple Silicon Mac image 860, and the Intel Mac image 244. Linux packages were in the dozens to low hundreds each. So the audience is phones and Windows first, which fits the CapCut framing better than the developer one.

The offline models are real, and they are not small

This is the part I would have guessed wrong. "Local Whisper captions" sounds like one download. The system requirements page lists the model families and their sizes: Whisper in tiny, base or small, each in English or multilingual, at 78, 148 or 488 MB; Kokoro text-to-speech at 132 or 350 MB; Pocket TTS voice cloning at 98 MB and Chatterbox cloning at about 1.1 GB across nine files; cutout models at 15, 179 and 40 MB; and a 5 MB enhance model. The app itself is about 500 MB with the smallest caption model, and about 3.4 GB if you take everything.

The floor is low. Any 64-bit CPU from 2013 or later, 4 GB of RAM, and no GPU at all: without one, the window and preview fall back to the CPU. The recommended machine is six cores, 16 GB for 4K timelines and the larger caption models, and any GPU with Metal, DirectX 12 or Vulkan. The same page says Windows on x64 and ARM64, macOS on Apple Silicon and Intel, Linux on x86_64 and ARM64, and Android are supported, with iOS in beta. It also says "14 languages," while the README says 25. One of those is stale; I could not tell which from the docs alone.

Two warnings the README gives up front. The Mac build is unsigned, so macOS will refuse to open it until you run xattr -dr com.apple.quarantine /Applications/Concat.app. The Windows build is unsigned too, until the project's application to SignPath Foundation's free code signing is accepted; a code-signing policy page exists for exactly that reason. Neither is a red flag for a two-month-old project. Both are a reason not to hand it to a non-technical client this week.

The API is shipped. The MCP server is not

The line that made me read the docs is in the README's highlights: "Scriptable. JSON-RPC, gRPC and MCP API, plus a CLI. AI agents can cut video with it." The TikTok did not mention this; the developer docs are where it gets precise, and they are unusually honest.

JSON-RPC is real and shipped. From a source checkout, cargo run -p concat-cli -- api '{"method":"version"}' answers with the API version, the Concat version and a capabilities list. The docs' quick start is five JSON lines piped into the same command: create a project, import a clip, add it at the first free slot, add a text clip at second one, and run an export. The export returns a job name immediately and then reports export.progress and export.done as events. Projects are folders, edits are commands with the window's own vocabulary and the window's own refusals, one export runs at a time, and errors carry codes you can branch on. That is a usable surface for an agent today, even without MCP.

gRPC is shipped behind a build feature. And MCP, per the page titled MCP in the developer docs: "An MCP server for Concat is planned and prototyped, not shipped." Checked against 0.2.6, the workspace has no MCP crate and no MCP dependency, and the docs say outright that the README's mention "describes the intent, not shipped code." There was a prototype in a pull request, closed and not merged because it re-implemented the transport instead of reusing the socket server. The feature request was closed in September with a pointer to a documentation issue. The design direction is to build it on the official Rust SDK as a third transport, which is the right call and also not a date.

The README says agents can cut video with it over MCP. The docs say the MCP server does not exist. Both are true: the API exists, the adapter does not.

What the docs give you instead is a bridge you write yourself: run concat-cli serve on localhost with a token, then a stdio MCP server in Python, about forty lines with FastMCP, that turns tool calls into JSON-RPC calls. They sketch the tools: version, open and create project, import media, apply an edit, list the effect catalogue, run an export and wait for the job. They even include a numbered list titled "Guidance for agents" that tells a model to call version first, keep the project folder path, read returned clip ids rather than guess them, batch related edits, and never start a second export while one is running. I have seen commercial products with worse agent docs than this unshipped feature has.

Licensing changed once, and it matters if you build on it

Concat started under MPL-2.0 and relicensed to AGPL-3.0-or-later on September 2. The commit message explains why: MPL is file-level copyleft, which would let a competitor absorb the engine into a closed product, and the binaries already statically linked a GPL component anyway. On October 12 the project went further and removed the commercial licence option entirely: AGPL and nothing else, with a plugin exception so that a plugin reaching Concat only through its API keeps its own licence. The developer docs restate the rule for API users in one sentence: a program that talks to Concat over the API is a client.

If you are cutting your own videos, none of that touches you. If you are thinking of wrapping Concat in a hosted service, the network clause of AGPL does, and the project chose it on purpose. The repository also carries a security policy, a DMCA address, a trademark note, and one paid sponsor slot, currently a proxy company. The commit history is dense with agent-authored commits, which I note without judgment; the audit folder and the relicensing rationale read like a maintainer who knows what they are doing.

What I would do with it, and what I have not done

The honest status: I have read the README, the requirements page, the developer docs, the MCP page, the licence history and the release assets. I have not installed the Mac build, not run the quarantine command, not fed it a clip, and not timed a caption pass. Everything about quality, stability on an 8 GB Mac, and how good the Whisper small model is on non-English speech is still a question. I run two brands, one in Arabic, and multilingual captions on a 488 MB model are the test that decides whether this replaces anything in my week.

Here is the test plan I will run, so you can run it before I do. Install 0.2.7, clear the quarantine flag, download only the multilingual small Whisper model. Drop in one three-minute talking-head clip and ask for captions; count the corrections. Then skip the GUI, start concat-cli serve, and script the five-line edit from the docs against that same clip, exporting vertical. If the captions need fewer than ten fixes and the scripted export runs clean, I write the forty-line MCP bridge and hand the whole thing to Claude Code as a tool. If either fails, Concat stays a GUI editor on the shelf next to the others, and that is still free.

The thing worth taking from this even if you never install it: a two-month-old editor with 5,000 stars has a cleaner contract for agents than most of the SaaS I pay for, because it treated the API as the product and the window as one client of it. The MCP server will arrive when someone writes the adapter. The hard part is already done.

Sources

Facts and prices checked October 15, 2026: