top of page

We Ran FFmpeg Skill: The Local Video Editor for Your AI Coding Agent

2 minutes ago
7 min read

The video uses a digital avatar of Erick Grau, CEO of Chibitek.

Somebody on your team, or a client's marketing team, is going to ask an AI coding agent to cut a clip, add captions and export it for social. That used to mean a video editor and a person who knows it. Now there is a project on GitHub that gives the agent its own editor, on the local machine. It is called FFmpeg Skill, by the developer kajisho5, and it has been climbing fast. Before anyone installs it on a work laptop, here is what we saw.

What Did We Actually Do?

On Saturday, Nov. 21, 2026, at about 6:32 AM ET, we installed FFmpeg Skill version 4.4.0 from npm in a locked down sandbox. That matches the GitHub main branch at commit 636e749. We gave the sandbox its own empty home folder with three fake AI tool folders (for Claude Code, Cursor and Codex), and we recorded every network connection.

Inside that sandbox we:

  • Ran the one-command setup, npx ffmpeg-skill setup, on a machine with no FFmpeg.

  • Ran four editing tools (cut, reframe, the YouTube Shorts export and a pixelate redaction) on test clips we generated ourselves.

  • Drove the project's MCP server with a small script.

  • Ran the uninstall.

We did not run real sudo, apt-get, Homebrew or winget. We used a stand-in for sudo that only logged what it was asked to do. We did not run transcription, use real phone footage, or open a real Claude Code, Cursor or Codex session. We used no real accounts. Everything below comes from our own run logs or the public repo, and where we say "in our run," that is what we saw in our sandbox at that commit on that date.

What Is FFmpeg Skill?

In plain words, it gives AI coding tools a local video editor. The GitHub description calls it a "Local video editor for coding agents: 42 FFmpeg tools with a machine-readable contract and an MCP server, for Claude Code, Cursor and Codex." Your agent can cut, caption, reframe, adjust loudness and export video on your own machine.

Under the hood, it is a set of Python tools that drive FFmpeg, plus an Agent Skill file and an MCP server. The README says you need Node 16+ and Python 3.10+, and the installer asks for FFmpeg 6.1 or newer.

Repo facts, dated:

  • Stars: 9,070 and 827 forks on the GitHub API at about 6:35 AM ET on Nov. 21, 2026, when we checked it for the video. At 1:39 PM ET the same day, the GitHub API showed 9,188 stars and 834 forks.

  • License: MIT, per the GitHub API, the repo's license file and the npm registry.

  • Version we tested: 4.4.0, commit 636e749, committed Nov. 21, 2026, at 6:25 AM ET.

  • Since then: by 1:39 PM ET the same day, main had moved to commit 8c85417 and npm had 4.4.1, 4.4.2 and 4.5.0. We rechecked the code behind each fine print item below on main, and it reads the same.

What Was the Good News?

Good news first: no calls beyond the npm download, npm copy matched the GitHub code, refused to overwrite a file, refused to write over its own source, no hidden prompts aimed at AI.

Credit where it is due:

  • No network calls beyond the npm download, in our run. The install and uninstall only reached the npm registry. During the editing tools and the MCP server, our logs recorded zero outbound connections. The project's PRIVACY.md says the same: "The tools send no telemetry, analytics or usage data."

  • The npm copy matched the GitHub code. In our run, all 86 installed files matched commit 636e749, with zero differences.

  • Guard rails on files. In our run it refused to overwrite an existing file without the overwrite flag, refused to write over its own source, and refused to send output to a web address.

  • No hidden prompts aimed at AI. We scanned the README, the skill file and the docs and found none.

  • Clean uninstall of the copies it installed, plus releases published to npm with provenance, and CodeQL and Scorecard checks in the repo's workflows.

We found nothing malicious in what we scanned and ran. That is not a guarantee, but it deserves credit.

What Is the Fine Print?

These come from our own run and the project's own code, not our guesses.

Fine print 1: a test clip with a map location tag kept that tag after cut, reframe, YouTube Shorts export and pixelate redaction. Apple style tags did not carry over.

1. Location tags stay in your video after an edit. We tagged a test clip with a made up map location. In our run, after a cut, a 9:16 reframe, the YouTube Shorts export and even a pixelate redaction, every output still carried that location tag. That is normal FFmpeg behavior: the FFmpeg documentation says "By default, global metadata is copied from the first input file." We found no option in these tools that strips it. One nuance: Apple style location tags in our test file did not carry over. The standard tag did. So do not assume either way. Check the file.

Fine print 2: the setup command installed into Claude Code, Cursor and Codex folders, then ran sudo apt-get install -y ffmpeg. No question asked. Our stand-in sudo logged it; nothing was installed.

2. The setup command on the install button runs a package manager without asking. The README's install button points at npx ffmpeg-skill setup. The README says it "installs the skill for every agent it finds on the machine," and installs FFmpeg when FFmpeg does not run, with brew, winget or sudo apt-get install -y ffmpeg, "printed before it runs." In our run, on a test machine with no FFmpeg, it installed into all three AI tool folders it found, printed the line, then ran sudo apt-get install -y ffmpeg. No yes or no question. We used a stand-in for sudo, so nothing real was installed. It is disclosed, and there is an opt-out: --no-ffmpeg. On macOS and Windows we read the code for the Homebrew and winget paths but did not run them.

Fine print 3: the MCP server lists 12 tools, but an unlisted tool was callable by name and wrote a file into the Claude skills folder. Output can land anywhere your account can write.

3. The MCP server lists 12 tools, but more are callable. The server's own code says it advertises "only the core 12" and that "Every tool is still callable by name." In our run, the tool list returned 12. We called one that is not on the list, by name, and it worked. It wrote its output into our Claude skills folder. We also sent a cut to a file outside the project folder, and that worked too. Output can go anywhere your account can write. A web address as output was refused.

4. It ships fast. On the morning we checked, npm showed 249 versions since the package was created on Sept. 3, 2026, including 4.4.0, published that morning. By 1:39 PM ET the same day, npm showed 252. The README documents how to pin an exact version.

What Should You Do If You Try It?

  1. Install FFmpeg yourself, and use --no-ffmpeg on setup. Then nothing runs a package manager on your behalf.

  2. Before you post a clip, check it for location tags and remove them.

  3. Check where every output lands. Point the agent at a working folder and look at what it wrote.

  4. Pin a version, and read the changelog before you update.

  5. Treat redact as a blur on pixels, not a privacy scrub of the whole file. The tool's own description says it blurs or pixelates a pixel rectangle you give it, and our run showed the location tag survives.

These five rules are Chibitek guidance, based on our sandbox run and the project's own code and docs. An AI agent that can edit video is still software with access to your files, and it belongs on the same list as everything else you manage.

Key Points of FFmpeg Skill

  • FFmpeg Skill by kajisho5 is an MIT licensed local video editor for AI coding agents, with an MCP server.

  • Chibitek ran version 4.4.0 (commit 636e749) in a sandbox on Nov. 21, 2026, with test clips, fake agent folders and a stand-in for sudo. Verdict: PASS WITH NITS. Not malicious.

  • In our run, edited and exported files kept the source's standard location tag, even after a pixelate redaction.

  • The setup command installed into every AI tool folder it found, then ran sudo apt-get install -y ffmpeg with no question asked. --no-ffmpeg turns that off.

  • The MCP server lists 12 tools, but unlisted ones are callable by name, and output can land outside the project.

Frequently Asked Questions

Did you run it with a real AI agent? No. We drove the MCP server with a script, in a sandbox, with fake agent folders. We did not open a real Claude Code, Cursor or Codex session.

Did setup really install FFmpeg with sudo? Setup called sudo apt-get with the yes flag, but our sudo was a stand-in that only logged the command. Nothing was installed. The code and the README both show that line.

Is FFmpeg Skill open source? Yes, under the MIT license.

Does it send my video anywhere? In our run, the editing tools made no network connections. The project's privacy statement says nothing is uploaded. Transcription is the one exception the project lists: it uses a speech engine you install, which may download its own model. We did not run transcription.

Is the location tag issue a bug? No. It is how FFmpeg copies metadata by default. It only becomes a problem if someone posts a clip without checking it.

Let's Review Before You Install

At Chibitek, we look closely at tools like this before they get anywhere near a client machine, with no geek-speak. If your team wants AI agents editing video, contact us at hello@chibitek.com and we can talk through where it should run and who should own it.

Install FFmpeg yourself. Use --no-ffmpeg on setup.

We make IT effortless so you can create, disrupt, and grow!

Follow Chibitek on YouTube at @chibitekmsp, and follow Chibitek on LinkedIn.

Sources

Every repo claim was checked against the file at the tested commit (636e749bd98fbd30c808f9aa55092ae388900936, committed Nov. 21, 2026, 6:25 AM ET), fetched from GitHub on Nov. 21, 2026, about 1:40 PM ET, then against the current main branch (8c85417, npm 4.5.0) and Chibitek's own sandbox run logs from Nov. 21, 2026.

  1. FFmpeg Skill repository

  2. README at the tested commit (install button, setup behavior, requirements, version pinning)

  3. Installer at the tested commit (agent folders, package manager commands, --no-ffmpeg)

  4. MCP server at the tested commit (12 listed tools, every tool callable by name)

  5. Metadata tool at the tested commit and redact tool

  6. PRIVACY.md at the tested commit (no telemetry, transcription exception)

  7. Release workflow at the tested commit (npm publish with provenance)

  8. License file (MIT)

  9. GitHub repository API (stars, forks, license, description; 9,070 stars on Nov. 21, 2026, about 6:35 AM ET, and 9,188 at 1:39 PM ET)

  10. npm registry metadata for ffmpeg-skill (versions, publish times, license) and the npm package page

  11. FFmpeg documentation (global metadata is copied by default)

  12. The video on YouTube and the LinkedIn post

Comments


bottom of page