Can a coding agent like Claude Code really edit video?
Yes, and a lot of people are doing it now. A coding agent can run commands, write files and call tools, and that covers most of what an edit is. In early November, Kino AI's Launch HN post described a desktop editor where "Codex / Claude Code can work on the same timeline" as your colleagues. Remotion now has an official page for coding agents. And guides like Learn With Hasan's walk through a full Claude Code pipeline built on ffmpeg.
What most guides don't tell you is that these are three different methods, and they give you three different things at the end. Pick the wrong one and you either get a file you can't fix or a setup that takes longer than the edit.
This guide compares all three fairly, then gives the exact commands for the one we build: connecting your agent to Takejam, a browser video editor, so it edits a real timeline while you watch.
What are the three ways, and how do they compare?
| ffmpeg scripts | Remotion code | A timeline over MCP | |
|---|---|---|---|
| What the agent does | Writes and runs commands that cut, join and encode | Writes React components that render to video | Calls an editor's tools: read the timeline, edit, look at frames |
| What you get | A finished MP4 | A code project plus a rendered MP4 | A normal editor project with clips, titles and keyframes |
| Fixing one small thing | Change the command and encode again | Change the code and render again | Drag it in the editor, or ask the agent |
| Can a non-coder take over? | No | No | Yes |
| Setup | Install ffmpeg (often Python and a transcription key too) | Node.js plus a Remotion project | One command, then sign in |
| Best for | Batch jobs: the same cut on 200 files, silence removal, format conversion | Motion graphics and videos made from data | One-off edits of footage you shot, and anything a person or team will review |
| Runs with no screen | Yes | Yes | Edits yes; checking frames and exporting need the project open in a browser |
Some honest notes on each:
- ffmpeg is free, fast and scriptable, and nothing beats it for repeating one job across many files. But the result is a flat file. If the third cut is half a second late, you edit a command and re-encode. The pipeline in the guide linked above also needs Python, Node, a transcription API key and a Claude plan, and its author says he has tested it on Windows 11 only.
- Remotion is excellent when the video is code: animated charts, templated videos, title sequences. Setup is
npx create-video --yes --blank my-video, thennpx remotion skills add, per its docs. It's less natural for cutting an hour of interview footage. Also check the licence: Remotion's licence page says it is free for individuals and companies of up to 3 people, and larger companies need a paid licence. - A timeline over MCP means the agent doesn't produce a file at all. It edits a project inside an editor, the same way you would, and you can keep going by hand. If MCP is new to you, read what MCP is, explained simply.
The rest of this guide covers the third method.
How do you connect Claude Code to a video timeline?
You need a free Takejam account, Chrome or Edge on a computer, and your clips on that computer.
- Open app.takejam.com, create a project and drag your clips and music into the Media tab. The files stay on your computer; they aren't uploaded.
- In a terminal, run:
claude mcp add --transport http takejam https://api.takejam.com/mcp
- Start Claude Code, type
/mcp, pick takejam and choose to sign in. (Runningclaude mcp login takejamin your shell does the same.) - A Takejam page opens in your browser. Sign in, pick the workspace the agent may edit and click Allow.
- Keep the project open in a browser tab and ask: "List my Takejam projects, open the newest one and cut it to 30 seconds."
By default, Claude Code saves the server for the current folder only. Add --scope user to the command to have it in every folder (Claude Code MCP docs).
How do you connect Codex or Cursor?
The address is the same. Only the place you put it changes.
Codex. Add the server by address, then sign in:
codex mcp add takejam --url https://api.takejam.com/mcp
codex mcp login takejam
OpenAI's Codex MCP page says the Codex CLI, the IDE extension and the ChatGPT desktop app share this configuration, so you set it up once. Flags change between Codex versions. If the add command is refused, put the server in ~/.codex/config.toml instead, then run the login command:
[mcp_servers.takejam]
url = "https://api.takejam.com/mcp"
Cursor. Create .cursor/mcp.json in your project, or ~/.cursor/mcp.json to have it everywhere (Cursor's MCP docs):
{
"mcpServers": {
"takejam": { "url": "https://api.takejam.com/mcp" }
}
}
Cursor opens the Takejam sign-in the first time it uses the server.
Anything else. Any app that accepts a remote (HTTP) MCP server works the same way: VS Code, Windsurf and others. In Takejam, Connect your AI → Other apps has the address and the JSON ready to copy.
Running the agent somewhere a browser sign-in can't open, such as an SSH session or a script? Make a key under Workspace settings → AI connector keys in Takejam and send it as the header Authorization: Bearer tj_…. Keep it in an environment variable, not in a file you commit. Only the workspace's owner and admins can create keys.
What can the agent actually do once connected?
It gets eight tools. The ones that matter most:
get_projectreads the timeline: tracks, clips with their times in seconds, the media in the bin, markers, scenes and the music's beat grid.edit_projectapplies a list of steps as one change. Steps include adding, trimming, splitting, moving and deleting clips, speed and speed ramps, keyframe animation, effects, transitions, templates, aspect ratio, markers, caption styles and snapping cuts to the beat. If any step fails, nothing changes and the agent is told why.get_referencelists the real effect, look, font, transition and animation names, so the agent doesn't have to guess them.look_at_framesrenders up to six moments, or a contact sheet of up to 24 frames, exactly as the export would, and returns them as images.undo_last_ai_editputs the project back to how it was before the last AI edit.
The other three list your projects and read or answer your team's review notes.
That frame tool is what makes an agent good at this. Coding agents already work in a loop: change something, run the tests, fix what failed. Here the loop is edit, look at the frames, fix what looks wrong.
What should you ask it?
Talk to it like an editor, with a length and a feel in mind:
- "Read my Takejam project 'Launch day'. Cut the clips into a 30-second 9:16 reel, strongest shot first."
- "Snap the cuts on the main video track to every 2 beats of the music." (Run Detect beat on the music first; see editing to the beat.)
- "Add a bold title in the first 2 seconds and a lower third with the speaker's name at 0:08."
- "Look at frames at 1, 8 and 20 seconds. If any text covers a face, move it."
- "Make a 16:9 version for YouTube and check that nothing is cropped badly."
One thing a coding agent can do that a chat app can't: read your repo. Point it at your README, changelog or launch notes and say "use the three features in CHANGELOG.md as the on-screen titles". The copy comes from your real project, not from the agent's guess.
If an edit goes wrong, click Undo AI change at the top of the editor, or tell the agent to undo its last edit.
What can't it do?
- It can't import your files. You drag media into the project yourself. The agent can only use what is already in the bin.
- It doesn't watch or hear your footage. It reads the timeline as data and sees a few rendered frames. Good file names help a lot: "demo-dashboard.mp4" tells it far more than "IMG_4031.MOV".
- It can't export. You click Export video… in the editor. There is no headless render, so this isn't the right tool for unattended batch jobs. Use ffmpeg for those.
- Looking at frames needs the project open in Chrome or Edge on someone's computer, because that is where your footage is.
- It doesn't generate footage, and it leaves scenes alone if a reviewer has marked them approved.
- The Free plan allows 50 AI edits a day. Each
edit_projectcall can hold many steps, so that goes further than it sounds. - No phones. Takejam runs in Chrome or Edge on Windows, Mac and Linux.
Which method should you pick?
- The same job on many files, or a pipeline that runs by itself: ffmpeg scripts.
- A video that is mostly animation, or one built from data: Remotion.
- Footage you shot that needs cutting, titles, music and a human eye at the end: a timeline over MCP.
They also combine well. Let the agent strip silence from a long recording with ffmpeg, drop the result into Takejam, then have the same agent cut and title it on the timeline.
Takejam's Free plan covers everything in this guide: the full editor, 1080p export with no watermark and 50 AI edits a day. Creator is $5 a month ($48 a year) for unlimited AI edits and 4K. Team is $8 per seat a month (minimum two seats), and adds live co-editing, so teammates watch the agent's edits land and can step in. Create a free project and try it with one command.
Prefer a chat app to a terminal? See the guides for Claude and ChatGPT.
Questions, answered
Can Claude Code edit videos?
Yes, in three ways. It can run ffmpeg commands that output a finished file, write Remotion code that renders one, or connect to a video editor over MCP and edit a timeline. The last one leaves you with a project you can still change by hand.
What is the command to add a video editor to Claude Code?
For Takejam it is claude mcp add --transport http takejam https://api.takejam.com/mcp. Then type /mcp inside Claude Code and sign in. A browser page asks which workspace the agent may edit.
Does this work with Codex and Cursor?
Yes. Codex adds the server with codex mcp add takejam --url https://api.takejam.com/mcp, then codex mcp login takejam. Cursor reads it from .cursor/mcp.json. All of them use the same tools, so the same requests work.
Do I need ffmpeg installed?
Not for the MCP method. The editor runs in your browser and does the rendering there. You only need ffmpeg if you choose the script method, or want to batch-process files before editing.
Are my video files uploaded?
No. In Takejam your media stays on your computer. The agent reads the timeline (which clip is where, titles, effects) and, when it checks its work, a few frames rendered by your open editor. More in editing video without uploading.
Is it free?
Takejam's Free plan includes the connector, 50 AI edits a day and 1080p export with no watermark. Your coding agent is billed separately by Anthropic, OpenAI or Cursor.