faceless-agent — a Claude Code plugin that turns your script and voice into a documentary

You write the story and record it in your own voice. faceless-agent does the editing: it times an AI-generated image to every scene, adds slow camera moves, glitch cuts, a CRT look and synced subtitles, mixes music and sound effects under your voice, and builds the matching vertical short.

You never open a video editor. You talk to Claude Code, and it runs the pipeline.

The clip on the right is the plugin's own output: 49 seconds, rendered in 47 seconds on an M3 Pro.

🇪🇸 ¿Hablás español? — resumen rápido

faceless-agent convierte tu guion y tu voz grabada en un documental sin cara: una imagen generada por IA por escena, movimiento de cámara, cortes con glitch, look de tele vieja, subtítulos sincronizados, música y efectos, y además el short vertical del mismo caso.

Vos escribís y narrás. La IA sólo produce. No abrís ningún editor de video: le hablás a Claude Code.

Las tres skills están escritas en español: setup (prepara todo y te explica qué cuesta cada cosa), brand-setup (el look de tu canal) y nuevo-documental (el video, paso a paso).

Instalación:

/plugin install faceless-agent --marketplace MartinOlivero/faceless-agent

Probado de punta a punta sólo en macOS. El método completo, con las clases y los casos, se enseña en la comunidad de IamAutom.

Hecho por Martín Olivero / IamAutom.

What goes in, what comes out

You bring	The plugin makes
A script, split into scenes	One AI image per scene, in a consistent style
Your recorded voice (.wav or .mp3)	Word-level transcript, used to place every subtitle and cut
A few timing decisions, in a small config.json	A 16:9 documentary and a 9:16 short, both mixed to −14 LUFS

The script and the voice stay yours on purpose: that is what makes the channel recognisably yours. It is not what keeps you within platform rules. YouTube looks at the picture, not the voice: realistic AI-generated scenes have to be declared as altered or synthetic content when you upload, and what gets penalised is failing to declare them, not using AI.

Install

/plugin install faceless-agent --marketplace MartinOlivero/faceless-agent

On Claude Code older than 2.1.275, use the two-step form:

/plugin marketplace add MartinOlivero/faceless-agent
/plugin install faceless-agent

Use

1. Set up once. Ask for the setup skill ("configurá el plugin"). It detects your system, offers to install the free tools, and walks you through the paid API keys, saying what each one is for and how it is billed before you sign up for anything.
2. Make it yours. Ask for brand-setup ("configurá mi marca"). It interviews you and writes ~/.faceless/brand.json: accent color, fonts, subtitle size, how much CRT, glitch and vignette you want (none is a valid answer), your outro and your short's call to action. Skip this and your videos carry a placeholder outro that says "TU NOMBRE".
3. Make a documentary. Ask for nuevo-documental ("armemos un caso"). It walks the whole pipeline: script, voice mastering, transcription, images, assembly, render, audio mix, the vertical short and the cover.

The one file you edit per video

Everything that changes from one video to the next lives in a config.json next to your images. Claude writes it with you from the transcript; you can also edit it by hand.

{
  "ESCENAS_INICIO": [0.0, 24.0, 46.0, 73.0],
  "KEN_BURNS": [[[1.15,0,0],[1.32,-70,-40]], [[1.30,-60,30],[1.12,40,-20]],
                [[1.12,50,30],[1.30,-50,-30]], [[1.32,0,-45],[1.14,0,35]]],
  "GLITCH_EXTRA": [1],
  "CALLOUTS": [[48.0, "NATANZ · IRÁN — 2010"]],
  "MUSICA": [["acto1.mp3", 0.0, 80.0, 0.18], ["acto2.mp3", 73.0, 999.0, 0.20]],
  "SFX": [["riser.mp3", 0.3, 0.40], ["impact.mp3", 46.0, 0.55]]
}

ESCENAS_INICIO is the second each image enters. KEN_BURNS is the camera move per scene, as a start and end [zoom, x, y]. CALLOUTS are on-screen fact plates. MUSICA and SFX name files and when they play. A misspelled key stops the run with an error instead of quietly falling back to the demo values.

Your channel's look is a second file, brand.json, written once. Both are passed in order, look first and timing second, so one video can override the brand when it needs to.

What it costs to run

The plugin is free and MIT-licensed. Running it uses:

• Free tools: ffmpeg, Node 22 or newer, uv, and HyperFrames, the engine that renders the video.
• Image generation, paid per use: a Gemini API key with billing enabled. Gemini has no free tier for images, so this one is required.
• Music and sound effects: HeyGen through HyperFrames (npx hyperframes auth login). Check their current free tier before assuming a cost.
• Optional: kie.ai, for two or three generated video shots when a scene needs real motion.

setup covers each of these one at a time.

What it is not

• Not a one-click channel generator. It does not write your script or clone a voice. If you want fully automated videos, this is the wrong tool.
• Not tested beyond macOS. The long form and the short have been run end to end on macOS only. CI runs the config tests and the dependency check on Linux and Windows, which is not the same as rendering a video there. Issues are welcome.
• Not in English yet. The skills are written in Spanish. Claude still answers in your language, and transcription takes a language code (see nuevo-documental, step 3).
• Not a YouTube uploader. It leaves the .mp4 files and the cover ready; you upload them.

Where the method comes from

The pipeline was built for Día Cero, a series of real hacking stories on the Martin Olivero channel. The published Stuxnet episode uses this same method, extended with archive footage that the plugin does not handle yet.

The plugin gives you the tools. How to pick a case, write it so people stay, and run a channel with it is taught in the IamAutom community (in Spanish).

What's inside

• skills/setup, skills/brand-setup, skills/nuevo-documental: the three skills.
• scripts/doctor.py: dependency check that prints JSON, used by setup.
• scripts/transcribir.py: voice to transcript.json with word-level timings (faster-whisper).
• scripts/armar_video_hf.py: builds the HyperFrames page for the 16:9 documentary.
• scripts/armar_short_hf.py: builds the 9:16 short from the hook.
• scripts/mezclar_audio.py: mixes voice, music and effects over the silent render.
• scripts/config_comun.py: the shared list of valid config keys.

The assembly scripts carry demo values from the Stuxnet pilot as defaults. Your config.json replaces them; the pilot's images and audio are not included.

License

MIT © Martín Olivero / IamAutom