Personal AI · Early alpha

Arynwood Cutroom

An AI-assisted video editor built on Kdenlive. Describe an edit in plain language and the Cutroom Assistant carries it out on your timeline, with a model on your own machine, asking before it changes anything.

Download 0.1.0 (Flatpak) Cutroom on GitHub
  • GPL-3.0
  • Linux
  • Flatpak 0.1.0
  • Early alpha

What Cutroom adds

Everything Kdenlive does, plus an assistant

Status: early alpha, Linux only. The 0.1.0 editor installs as a Flatpak. The assistant needs the mcp-kdenlive tool server, which is not published yet, so it can’t run from the download.

  • The Cutroom Assistant

    A chat panel docked in the editor window. Ask for a timeline summary or an edit, and it calls the editor’s tools to do it.

  • D-Bus scripting API

    Around 190 scriptable methods on the editor, from D-Ogi’s Kdenlive fork, for driving it from Python, the command line or any D-Bus client.

  • An MCP tool server, coming

    183 tools an AI agent can call: import, cut, arrange, transitions, effects, keyframes and render. Not published yet; the assistant panel needs it.

  • More Creative Commons media

    Jamendo music and Openverse, on top of Kdenlive’s own media providers.

  • Expressions and GPU effects

    A JavaScript expression engine for keyframes and two libplacebo GPU effects, from D-Ogi’s fork.

  • Installs beside Kdenlive

    The flatpak has its own application ID, com.arynwood.Cutroom, so it can sit next to a regular Kdenlive install.

How the assistant works

It asks before it changes your project

Open it from the Launch Assistant button, or View, then Cutroom Assistant.

  • Your model, your machine. By default it talks to Ollama on 127.0.0.1 and nothing leaves the computer. With a remote OpenAI-compatible endpoint, the conversation (including clip and file names) goes to that service. It never uploads the media files.
  • Allow or Deny on every change. Reading, moving the playhead and making previews never ask. It makes at most 12 tool calls per reply and 40 per request, and won’t repeat an identical change after two runs.
  • Only the relevant tools. The full tool list is about 20,000 tokens, more than a small local model can use, so it sends the tools that fit the request plus a search function for the rest.
  • No promise of undo. Many timeline edits can be undone, but not every scripted operation can. Save a copy of anything important first.

Before you install

What it needs, and its limits

What you need running

  • The Cutroom editor (the flatpak build is the tested route)
  • Ollama, or another OpenAI-compatible endpoint, with a model that supports tool calling
  • The mcp-kdenlive tool server on 127.0.0.1:8420, a separate process that isn’t published yet

Limits: read this first

  • Small local models are unreliable at multi-step edits. On “add all the photos in the bin to the timeline”, three runs each: the best 14B model we tried got it right once, hermes3:8b and mistral never did, and gpt-oss:20b didn’t answer within four minutes on a 12 GB card
  • The panel supports Ollama and the OpenAI chat-completions dialect only
  • Linux only; no Windows or macOS builds

Arynwood Cutroom is a modified version of Kdenlive. It is not made or endorsed by the KDE project or the Kdenlive team. It is built on the D-Ogi/kdenlive fork, whose D-Bus scripting API it depends on.

Questions

Cutroom questions

Privacy, models and what happens to your project.

Is Cutroom free?

Yes. It’s GPL-3.0 open-source software, like the Kdenlive it’s built on. The source is on GitHub.

Does my video leave my computer?

Not by default. The assistant uses Ollama on your own machine. If you point it at a remote endpoint, the conversation and clip or file names go to that service, but never the media files.

Can I use Claude with it?

Claude Code and Claude Desktop can drive the editor through the MCP tool server, from outside the editor. The in-editor panel supports Ollama and OpenAI-compatible endpoints, not Anthropic’s API.

Which model works best?

Larger models do better. In testing, small local models were unreliable at multi-step edits, which is why the panel asks before every change. A larger local model or a hosted one will do better.

Keep reading

Related tools and docs

MCP & creative workflows

Want a Kdenlive workflow set up for you?

Start with one task, such as a repeatable edit, and we’ll agree on the integrations and limits before any work begins.

Email for a quote See setup services

Running AI on your own Linux machine? Join the Linux Local AI Community on Facebook to compare setups and ask questions.