What Is Concat? A Free Local-First Video Editor That Lets Claude Code Cut Video via MCP

Opening — Cutting a 60-Second Shorts Video Takes a Whole Evening

Tuesday, 9 p.m. You have just shot three short clips: a configuration tutorial and a reply to a comment. Still to do: trim the hesitant opening, add Vietnamese captions, export at 1080x1920, and publish.

You open CapCut, sign in, and drag the footage in — the app suggests syncing to the cloud. Auto-caption and background removal sit behind a paid plan, and a template leaves a watermark on export. One clip done, and it is nearly 10 p.m.

For a developer, the frustration runs deeper: you hand repetitive work to Claude Code, but a pure GUI editor has no API for an agent to trim a clip by timestamp. Concat answers exactly that pain: a free, open-source, local-first video editor with a built-in MCP server so an AI agent can cut video.

What Is Concat? A Local-First Video Editor as a Free CapCut Alternative

Concat is an open-source video editor that runs on your machine, for macOS, Windows, Linux, and Android. The jub0t/Concat repo has 16,421 stars, 2,044 forks, an AGPL-3.0 license, and its latest release is v0.2.9, dated 05/11/2026 — verified via the GitHub API on 10/11/2026.

Concat solves three problems: cost — no watermark, no account, no subscription, no upload; privacy — sensitive footage never leaves your machine, and AI models run offline after the first download; and automation — command-level control via a JSON-RPC API and MCP server.

It fits developers using Claude Code, creators making Reels/Shorts who need fast captions without a watermark, and IT staff who prefer local-first tools with no cloud. If you need CapCut templates, cloud team collaboration, or an absolutely stable editor for a commercial deadline, it is not yet the right choice.

Concat supports multi-track editing, keyframes, 100 effects, titles, and background removal. Auto-caption runs on Whisper on-device; TTS and voice cloning use Kokoro and Chatterbox, models downloaded once — Whisper 78–488 MB. Export goes up to 4K 60fps 10-bit with H.264, HEVC, and AV1.

How Concat Works — Rust Engine, GPU Compositor, and MCP Server

Concat has three layers: a GUI with a multi-track timeline and a built-in MCP server sharing one processing core; a Rust engine with a GPU compositor, which decides export performance and VRAM usage; FFmpeg 7+ for H.264, HEVC, and AV1; and on-device AI with Whisper, Kokoro, and Chatterbox.

The key point is the JSON-RPC API between the MCP server and the engine: it turns the timeline into a command-addressable state — create a project, trim a clip, add captions, export video — so the GUI and the agent operate on the same project state.

User flow: import media, edit on the timeline, add captions, preview, then export a local file. Agent flow: Claude Code calls the MCP server with concat --mcp, the server translates the request into JSON-RPC commands, and the agent uses preview_frame to inspect a frame before exporting. This closed loop is what makes Concat’s MCP genuinely useful.

There are three MCP modes: --mcp window (through the open window), --mcp headless (engine only, no GUI), and the default --mcp auto. Window mode is not live mode — the agent directly editing the project open in the window is only a plan, issue #337.

flowchart TD
    A["User / Claude Code"] --> B["concat --mcp<br/>window / headless / auto"]
    B --> C["JSON-RPC API"]
    C --> D["Rust engine + GPU compositor"]
    D --> E["Project state / Timeline"]
    D --> F["On-device AI<br/>Whisper + TTS"]
    G["Local footage"] --> D
    E --> H["preview_frame"]
    H --> A
    E --> I["FFmpeg export<br/>4K 60 10-bit"]
    subgraph LOCAL["Local-only: no upload, no account"]
        B
        C
        D
        E
        F
        G
        H
        I
    end
    B -. "writes only under home folder; CONCAT_MCP_ROOTS widens it" .-> E
    style LOCAL stroke-dasharray: 5 5

Building from source needs Rust 1.93+, FFmpeg 7+ with dev libraries, cmake, and a C++ compiler. Note the export-error reports on NVIDIA cards with 8 GB VRAM, issues #223 and #297 — I have not verified whether they are closed in v0.2.9.

How to Install Concat and Connect Claude Code via MCP

The version used here is v0.2.9, released 05/11/2026 — the latest release on GitHub, verified 10/11/2026. Note: the README badge shows 0.2.10, but there is no 0.2.10 release yet. Download from concat.standalones.org or GitHub Releases.

On macOS, download the Apple silicon or Intel build and drag it into Applications; the build is unsigned, so macOS may block it on first launch:

xattr -dr com.apple.quarantine /Applications/Concat.app

This removes the quarantine flag — use it only with a file from the official website or GitHub Releases.

On Windows, use the per-user installer at %LOCALAPPDATA%\Programs\Concat\concat.exe, or the MSI build. On Linux, choose .deb, .rpm, .AppImage, the Arch package, or Flatpak; the binary is at /usr/bin/concat. After installing, open Settings to download the AI models on first run — this needs network; later runs work offline.

Connecting Claude Code on macOS takes one command from the official MCP documentation:

claude mcp add concat -- /Applications/Concat.app/Contents/MacOS/concat --mcp

Binary paths: Windows uses concat.exe; Linux uses /usr/bin/concat or the .AppImage file itself.

{
  "mcpServers": {
    "concat": {
      "command": "/Applications/Concat.app/Contents/MacOS/concat",
      "args": ["--mcp"]
    }
  }
}

MCP tools, with exact names from the documentation: project_create, media_import, clip_trim, clips_split, captions_add, title_add, keyframe_set, effects_list, export_start/export_wait, preview_frame.

Minimal workflow: describe it in words — “cut the intro 0:00–0:08, add Vietnamese captions, export 1080x1920” — Claude Code creates a project, imports footage, trims, adds captions, then uses preview_frame to check before exporting.

After installing, verify: open the app and confirm v0.2.9; run claude mcp list to check concat is registered; create a sample project, import a short clip, trim a few seconds, and try an export.

Common errors: macOS blocking the unsigned app — fix with xattr above; Windows SmartScreen, because the build is not code-signed — a warning, not an error; AI models not yet downloaded, so captions or TTS do not run offline on first use; the agent blocked from writing outside home — widen it only with CONCAT_MCP_ROOTS.

Technical Review — Strengths, Weaknesses, Security, and Operations

Strengths: genuinely free (AGPL-3.0, no watermark, no account, no subscription, no upload); local-first (footage and models stay on your machine, suiting sensitive video); and scriptable via MCP, making cutting a repeatable workflow.

Weaknesses: still beta — the project itself says it still has “edges”. Missing: auto-reframe, object and face tracking, profanity detection. Risk is concentrated: one main maintainer, jub0t, with about 1,100 commits, and 264 open issues. Biggest limit: no live mode yet — agent and user cannot edit the same open project, so hand the whole edit pass to the agent, review it with preview_frame, then open the GUI to check and export.

By default, Concat is fairly safe: no upload, no account, and the agent writes only under the home folder. But CONCAT_MCP_ROOTS widens the agent’s write scope beyond home — set it narrowly per project, never across a whole drive. And CONCAT_API_ADDRESS and CONCAT_API_TOKEN allow remote control; open a remote server without protecting the token and the network, and someone else could control the editor on your machine — use it only on a trusted network. AGPL-3.0: free to use, but embedding it in a commercial product requires sharing the source.

This article has no independent performance measurements, so it claims no render speed or export time. What is grounded: proxies make scrubbing 4K footage smoother, and export is heavy on GPU and VRAM — cards with 8 GB VRAM already have error reports. Concat updates frequently — 9 releases in 10 weeks — so read the changelog before upgrading, back up the project and original footage, and keep the old installer: there is no built-in rollback.

Concat vs CapCut and Other Options in the Same Range

Compared on price, watermark, cloud, on-device captions and TTS, agent control, templates, stability, export, and privacy:

CapCut wins on templates, ready-made effects, cloud collaboration, stability, and community. Concat wins on being completely free, no watermark, no upload, open source, and MCP for agents — trading templates for control and automation is a fair deal. DaVinci Resolve, Kdenlive, and Shotcut are more mature; if you only need to cut one long video by hand, they are still safer.

Concat in Practice — 4 Concrete Scenarios

Creators — Reels, Shorts, YouTube: batch cutting to a template. Ten short clips, one format. The agent trims each by timestamp, adds Whisper captions, inserts titles, and exports at 1080x1920 — checking with preview_frame first.

Developers — Claude Code as a video-cutting assistant. “Cut the intro 0:00–0:08, add Vietnamese captions, export 4K” — Claude Code calls the MCP tools, checks its work with preview_frame, then exports. Live mode does not exist yet: the agent works on its own project.

Business and IT — internal video that never leaves the machine. Training videos and internal screen recordings should not go to the cloud. Concat runs locally; project and exports stay internal. The same local-first philosophy as VoiceStudio, free local voice cloning, Handy offline speech-to-text, and Hex, local voice dictation in Rust; see also Claude Code with AWS Bedrock AgentCore.

Homelab and DevOps — automating video from a pipeline. A homelab machine runs Concat headless via MCP to render periodic release clips. Desktop-tool automation, not a render farm — always check the output before publishing.

Conclusion — Who Should Use Concat, Who Should Not, and the Next Step

Use it now if you are a developer who wants an agent to cut video for you; a creator who needs a free editor with no watermark and no cloud; someone working with sensitive video who needs local-first; or someone who likes new open-source tools and accepts beta software. Not yet if you need CapCut templates, vertical auto-reframe, tracking, cloud team collaboration, or a commercial deadline that cannot tolerate an export error.

Windows ARM users, or those with an NVIDIA card with 8 GB VRAM, should check issues #223 and #297 against v0.2.9 first. Next step: download v0.2.9, try one short clip, connect MCP to Claude Code with one command, run the sample workflow, then use it for real work. In short: free, local-first, and agent-controllable.

Lessons Learned

1. Local-first is real value when footage is sensitive — but you must handle the hardware, AI models, and backups yourself.
2. MCP turns video cutting into a repeatable workflow — the value lies in preview_frame and verified export.
3. Without live mode, agent and user work on two different copies of a project — agree on the source of truth before publishing.
4. Nine releases in 10 weeks and a single main maintainer are two sides of one coin: fast updates, but upgrading mid-project needs a backup and the old installer.
5. With no independent benchmark, do not trust a feeling of speed — verify export on your own machine and footage before a deadline.

Verification Scope

Checked 10/11/2026 against the official README and MCP documentation of jub0t/Concat and the GitHub API (release v0.2.9, 05/11/2026). Verified: repo information, features, MCP commands and binary paths per OS, and supported platforms. Defaults: the agent writes only under the home folder; widening via CONCAT_MCP_ROOTS and remote control via CONCAT_API_ADDRESS/CONCAT_API_TOKEN are optional; live mode is planned, not yet available. Numbers here are configuration, not benchmarks: Whisper 78–488 MB and 4K 60fps 10-bit export are specifications from the documentation. Not tested: this article did not install or run Concat; export errors on NVIDIA cards with 8 GB VRAM and the system-requirements page need re-checking against v0.2.9; the README badge 0.2.10 could not be matched to a release.

References

• Official repo: https://github.com/jub0t/Concat
• MCP documentation: https://github.com/jub0t/Concat/blob/main/docs/MCP.md
• Website and downloads: https://concat.standalones.org
• System requirements: https://concat.standalones.org/guides/system-requirements
• Releases: https://github.com/jub0t/Concat/releases