Skip to content
Profile
E. Burgos
November 13, 2026By Esteban Burgos7 min read

Video Maker: I gave any AI agent a video studio

#Video Maker#MCP#Claude Code#Agentes de IA#Remotion

Making a short video for social media looks cheap until you actually do it. A thirty-second idea ends up as recording, transcribing, cutting silences, adding captions, picking music, taking care of audio and exporting in each platform's format. With Video Maker, a product of my own that I have been building, I automated that path. In November 2026 I added something that interested me more: letting any AI agent drive it. This article covers how it is put together and which technical decisions I found worth sharing.

The problem: short video without hours of editing

Producing video for Reels, TikTok or YouTube takes hours of editing and pushes you toward paying editors or subscriptions. The work repeats on every piece: transcribe, trim silences, place captions inside each platform's safe zones, mix music without covering the voice, render. I wanted to automate the whole pipeline, from the idea or the script to the final file, without losing control over what comes out.

What Video Maker is today

Video Maker is a local video studio with an AI producer. It has two paths:

  • Motion graphics: you write a script, add photos, and the studio builds animated scenes with your brand, voice and music.
  • Raw material editing: you drop in recordings and the studio trims silences, adds kinetic captions of one to three words, inserts keyword-driven B-roll and mixes the music with ducking.

It outputs vertical 9:16 or horizontal 16:9 MP4, with presets for Reels, TikTok, Shorts and YouTube. The AI producer is conversational: it proposes the video on a card and you adjust it by chat. It can run on a Claude Code subscription or on free OpenRouter models, so using it does not force you to pay for an API.

The heavy lifting runs on the user's Mac: transcription with whisper.cpp, voice with Piper and sherpa-onnx (with Rioplatense Spanish voices), mixing and rendering with FFmpeg. I said "local" and I want to be precise about the caveats: the text you give the AI producer goes to Claude or OpenRouter; if you enable ElevenLabs (optional, with your own API key), that text goes to their cloud; and publishing to YouTube uploads the video to the user's account. In those three cases it is not "all local", and I would rather say so.

The desktop app is open source (MIT) and can be downloaded for macOS 13 or later on Apple Silicon from https://video.estebanburgos.com.ar. It is not signed with an Apple Developer certificate, so the first launch needs an extra Gatekeeper step, explained in the install guide.

The jump to agents: plugin and MCP server

The question I asked myself was what would happen if you did not need to open the studio to ask for a video. Since November 2026 there are two public pieces:

  • A Claude Code plugin, video-maker@video-maker, with five skills. video-maker is the entry point: it reads the repository, picks the type of video, agrees it with you in one round, produces and reviews. video-maker-brand derives the brand from the repo: name, spoken name, palette with measured contrast, logo, fonts, voice and music. video-maker-motion-brief covers script, storyboard, render and visual review. video-maker-screen-capture records any web UI with Playwright at 1080p30, with a timed script. And video-maker-raw-material handles your own clips, photos, voice-over and music, plus free media with its licenses.
  • A public MCP server on npm, @e-burgos/video-maker-mcp, with about 40 tools for projects, brands, assets, motion, timeline, jobs, music and free media, voice and review.

The result is that you can tell an agent "make a tutorial of the checkout" or "a 45-second promo" on any repository (an API, a library, an app, a store) and it produces the video with the local studio. The MCP server discovers the studio on its own, whether it is the desktop app or pnpm studio in a clone of the repo; if needed, you can force it with environment variables.

One design decision I am happy with: publishing and deleting are not exposed to the agent. The person does those two actions from the studio. An agent can get a take or a text wrong, and that gets fixed in the next iteration. But a publication to a real account or a deletion is not undone by another iteration. I prefer the agent to do the work and the final word, the one with consequences outside the studio, to stay human.

Integration with SDD-Harness

Video Maker was built with Spec-Driven Development, the methodology I use in my repos: in main there are 24 registered specs, 20 complete and 4 in draft. That same discipline led me to integrate it into the SDD-Harness kit, so anyone working with the kit can add it with a single command:

bash
pnpm sdd:video -- --enable

The command declares the plugin in Claude Code and the MCP in Copilot, Gemini and Antigravity, with the version pinned. The idea is that the video studio becomes one more capability of the harness, available in whichever agent each person prefers, with no manual per-provider setup.

Technical decisions worth sharing

Two render engines. For raw material I use Remotion 4, which treats video as React components. For scene animation I use HyperFrames (HTML and GSAP). Each engine is the most natural one for its job.

Contracts in timeline-core. One Timeline feeds the preview, the render and the editor. When I added the MCP server, that decision paid off: the agent edits the same object the person sees, with the same types. A lib called studio-capabilities also describes what the studio can do (formats, presets, caption styles, scene types, voices) so the agent does not have to guess.

Careful audio. Masters go through a limiter and true peak is measured. Music always sits below the voice, with a bank per video type that rotates, plus free music from Openverse and Wikimedia Commons and CC0 or CC BY sound effects with automatic credits.

Variety by seed. Logo bumpers, transitions (wipe, flip, clock-wipe, iris), on-screen text and kinetic typography rotate with a seed per video. That way two videos from the same brand do not come out identical, but each one stays reproducible.

Explicit licenses. The code is MIT, but the app bundles third-party binaries, models and voices under their own licenses, listed on the Licenses page of the site. Remotion is not free software: it has its own license, free for individuals and small companies under its thresholds, and a company above them needs Remotion's company license. HyperFrames is Apache-2.0.

How to try it

You need the studio running on your machine (the desktop app open, or pnpm studio in a clone of the repo) and Node.js 22 or newer. Then, in Claude Code:

text
/plugin marketplace add e-burgos/video-maker
/plugin install video-maker@video-maker

If you prefer to use only the MCP server from another client:

bash
npx -y @e-burgos/video-maker-mcp

Recorded tutorials need Playwright with Chromium and ffmpeg. The app is at https://video.estebanburgos.com.ar and the code at https://github.com/e-burgos/video-maker.

What I take away

Three things. First: giving an agent tools with clear contracts pays off more than giving it freedom; the MCP works because the studio already had types and a single Timeline. Second: keeping the irreversible out of the agent's reach is a product decision, not a limitation. Third: documenting licenses and the caveats of "local" from the start avoids surprises later.

There is still room to improve: for now it is macOS on Apple Silicon only, and I have no verified usage metrics. If you want to exchange ideas about video pipelines, agents with their own tools or MCP, you can find me on LinkedIn.

One last fact: the video at the top of this article was produced by Video Maker itself, driven by a Claude Code agent with this plugin.

Written by Esteban Burgos

Comments

Be the first to comment.