Content Factory

Content Factory is a free, open-source faceless video generator for Apple silicon Macs,

and a local alternative to Higgsfield Content Factory. It turns a script into a 30–120 s vertical video

for TikTok, Shorts or Reels from about 8 AI images a minute, with voiceover, subtitles and music.

By Jutsu Studio, open-source Mac tools for AI video.

Quick start  ·  How it works  ·  vs Higgsfield  ·  vs MoneyPrinterTurbo  ·  Claude and MCP  ·  FAQ  ·  Architecture  ·  Contributing

At a glance

Compared with Higgsfield Content Factory and with
MoneyPrinterTurbo and short-video-maker: see below. What
Content Factory does not do: generate video with a video model, write the script itself, post for you, or run on
Windows, Linux or Intel Macs yet.

Demo

What if every human vanished today? 57 s · AI frames · Low-poly City 8 beats, a source for every fact. Script	What if the Sun went out? 31 s · Drawn · gouache Every scene is SVG code, no image generated. Script	What if you fell into Jupiter? 53 s · AI frames · Space Epic Cuts between a wide shot and punch-ins on the diver. Script

Claude wrote each script and handed it to the engine over MCP; the scripts are in formats/what-if/examples/. The engine
did the rest: frames, depth parallax, transitions, narration, word-by-word subtitles, CC0 music ducked under the voice, and a quality check.

Why Content Factory

Faceless short-video channels run on formats: the same structure, a new topic every day. Content Factory turns a
format into a pipeline you own.

• No video model. Motion comes from a handful of AI stills: depth parallax, frame interpolation, layers and
  camera moves. A one-minute video needs only eight images.
• Formats are data. A format is a folder with a script schema, an author's guide, prompts, style presets and a
  render template; the engine, the app and the MCP server read whichever formats are there. The bundled reference
  format is What if: one world, one axis of change, a counter on screen and a loop back to the hook.
• Edit one beat, rebuild one beat. Every stage is cached by the hash of its input, so a changed line re-voices
  one beat and a new take of a frame re-renders only what depends on it.
• A human at the gate. You write or approve the script and accept the result. Nothing is posted for you.
• Local first. Voice, speech alignment, depth, interpolation and rendering run on your machine. Only image
  prompts go to the provider you picked, with your own key.
• Agent-ready. An MCP server lets Claude Code or any MCP client write scripts, run the pipeline and pick up the
  result, sharing one queue with the app.

Open-source alternative to Higgsfield Content Factory

Higgsfield Content Factory turns a topic into a
faceless What if video in the cloud. Content Factory makes the same kind of video as an open-source app you run
yourself.

Content Factory is an independent project, not affiliated with or endorsed by Higgsfield AI. Higgsfield is a
trademark of its owner.

Compared with MoneyPrinterTurbo and short-video-maker

MoneyPrinterTurbo and
short-video-maker are the best-known open-source faceless video
generators. Both cut a voiceover over background footage, mostly stock video from Pexels. Content Factory draws every
beat instead: one AI still per beat in one style, so the same place, character and camera carry through the video
and only the state of the world changes.

Pick MoneyPrinterTurbo for footage-based videos on any OS, short-video-maker for an n8n or REST pipeline over stock
video, and Content Factory when the video is a story that needs its own consistent world, such as What if.

What it does

• Channels: a project per TikTok, YouTube or Instagram account keeps its videos alike. A channel has variants
  (ready settings for one kind of video: format, style, language, voice, length, hashtags) and a cast of recurring
  characters whose looks text and reference images go with every frame, so they look the same in every video.
• New video: format, topic, style, language, length (within the format's limits), narration or on-screen text
  only.
• Script editor: the format's fields per beat (for What if: counter, caption, narration and image state),
  checked against its schema.
• Frames: six style presets or your own; another take of any frame in one click.
• Drawn mode (prototype): pick Drawn in New video and Claude draws each beat as an SVG scene, animated by data
  attributes and painted with a gouache, ink or crayon look. No image generation, so a video costs nothing.
• Voice: one engine per language, set in profile.yaml: Kokoro on MLX by default, Vosk for Russian (an optional
  component); pauses
  ([pause 0.8]), loudness normalised to −14 LUFS.
• Subtitles word by word from forced alignment.
• Music from a local library of CC0 tracks, ducked under the voice.
• Quality check: length, black frames, loudness, speech clarity, text length, music license.
• Copies in other languages that reuse the frames: English, Spanish, German, French, Portuguese, Russian.
• Review and export: copy the video file to the clipboard, show it in Finder, save accepted videos to a folder.
• Queue with live logs and cancel; Storage settings with disk usage and cache cleanup.

Quick start

There are no signed builds yet, so build the app from source. You need macOS on Apple silicon,
uv, Node.js 20+, FFmpeg and
Rust.

brew install uv node ffmpeg
git clone https://github.com/jutsuhq/content-factory && cd content-factory
cd app && npm install && npx tauri build --bundles app
open "src-tauri/target/release/bundle/macos/Content Factory.app"

On first launch the onboarding:

1. Installs the components: the video renderer (Remotion, about 1 GB), RIFE for transitions, the Depth
   Anything V2 model and the starter music. FFmpeg and Node.js come with a one-line install command if missing.
2. Asks for an image provider key and checks it without generating anything. Drawn videos need no key, so this step can wait.
3. Sets your defaults: language, style, and narration or on-screen text only.

Everything can be changed later in Settings.

Image providers

Several providers can run at once and share the frames.

How it works

flowchart LR
  S["Script<br/><sub>you or Claude</sub>"] --> A{{"Approve"}}
  A --> F["Frames<br/><sub>image API</sub>"] --> D["Depth<br/><sub>Depth Anything</sub>"] --> T["Transitions<br/><sub>RIFE</sub>"] --> R
  A -. drawn mode .-> R
  A --> V["Voice<br/><sub>Kokoro</sub>"] --> W["Align<br/><sub>Whisper</sub>"] --> R["Render<br/><sub>Remotion</sub>"]
  R --> Q["Quality check"] --> X{{"Accept"}}

Every script names its format; the format decides the fields, the checks and the Remotion composition that
renders it. In AI-frames mode the first frame sets the style; every other frame is an edit of it that changes only
the state of the world, so the camera holds still while the world changes. See docs/architecture.md for the engine, the
format files and the job queue.

From the command line

The engine is the Python project in engine/; run these there.

cd engine
uv sync                                    # add --extra vosk for the Russian voice
uv run factory doctor                      # detect the machine, write profile.yaml
uv run factory setup all                   # renderer, RIFE, depth model, music
uv run factory catalog format              # formats and how to write a script in one
uv run factory create "<topic>" --script beats.json
uv run factory create "<topic>" --project <id> --variant <id> --cast milo,biscuit
uv run factory run <slug> approve          # make the video, wait for the job
uv run factory run <slug> rebuild --stages render
uv run factory accept <slug>

The command line is a client of the engine's API, like the app and the MCP server: the same commands as the MCP
tools, one shared job queue. uv run factory --help lists them all.

Claude and MCP

Content Factory is an MCP server for the whole AI video pipeline, from script to rendered video. factory mcp is an MCP server over stdio. It talks to the running engine (and starts it if needed), so the app and
the agent see the same queue. The factory never writes scripts itself: the agent lists the formats and reads one's
rules with catalog('format'), writes the script and passes it to create with the format id.

claude mcp add -s user content-factory -- uv run --project /path/to/content-factory/engine factory mcp

Tools: status, episodes, episode, create, edit_script, run (approve, rebuild, redo a frame, translate),
job, accept, catalog (skill, format, new_format, styles, music, voices, projects), format_source, save_format; prompt make-video. Then ask Claude something like "make a What if
video about falling into Jupiter, 70 seconds, space epic style", or "two new videos for Tiny Town Tales, English
variant, with Milo": catalog('projects') gives the channel's brief, variants and cast. When no format fits, Claude
can bring a new one: it reads the closest format with format_source, writes its own (schema, guide, prompts,
styles, Remotion template, an example) and save_format checks and adds it, ready in the app at once
(formats/README.md).

The MCP server also carries the make-video skill: how to make a video well, from the
idea test and the script rules to the checks on the result and a checklist before posting. Any MCP client gets it as
the make-video prompt and as catalog('skill'), which the server's instructions tell the agent to read before the
first video. The reasoning behind it, from picking an idea to staying original on a channel, is in
docs/playbook.md.

Repository layout

Each top-level folder is a module with its own README, so a change usually stays inside one of them.

Everything the engine writes in a checkout goes to data/, which git ignores.

Privacy

• No backend, no account, no telemetry. The engine listens on 127.0.0.1 only.
• Image prompts and the reference frame go to the image provider you chose, with your key. Nothing else leaves your
  machine.
• API keys live in the app's data folder (.env, owner-only permissions).
• Found a vulnerability? Report it privately as described in SECURITY.md.

Status

Early. macOS on Apple silicon is the tested platform. Windows and Linux builds are planned: the engine avoids shell
scripts and platform-only tools, but the voice back end there is not finished yet.

Credits

• Remotion renders the video. It is installed on your machine by the setup step and has
  its own license, which may require a company license for some teams.
• Depth Anything V2 small (Apache-2.0), via
  onnx-community.
• RIFE ncnn Vulkan (MIT) for frame interpolation.
• Kokoro (Apache-2.0) via MLX-Audio.
• Music: CC0 tracks from OpenGameArt, downloaded at setup.

FAQ

What is Content Factory?
An open-source desktop app and engine that turns a script into a finished faceless short video for TikTok, YouTube
Shorts and Instagram Reels: AI frames, parallax motion, transitions, voiceover, subtitles, music and a quality check.

Is it free?
Yes. Content Factory is free and open source under the MIT license, with no subscription, credits or watermark.
Drawn videos need nothing else; AI-frames videos draw their images with your own OpenAI, Gemini or fal.ai key.

Is it a Higgsfield alternative?
Yes, for faceless What if videos: the same format, made locally and open source. It does not generate video with
a video model; motion comes from stills.

Does it write the script for me?
No. You write it in the app, or Claude (or any MCP client) writes it through the MCP server. The engine checks it,
voices it and renders it.

Which platforms does it run on?
macOS on Apple silicon today. The engine avoids platform-only tools; Windows and Linux are planned.

Which languages can it narrate?
English out of the box with Kokoro, Russian with the optional Vosk voice; other languages take a voice set in
profile.yaml. Scripts and on-screen text can be English, Spanish, German, French, Portuguese or Russian, and a video
can be copied into another language that reuses its frames.

Can I add my own video format?
Yes. A format is a folder in formats/ with a schema, a guide, prompts, styles and a Remotion template;
see formats/README.md.

Related

Other open-source projects from Jutsu Studio:

• Jutsu Studio: the open-source Higgsfield desktop app for Mac · jutsustudio.com
• Swapjutsu: free character swap for Mac, no watermark · jutsustudio.com/swapjutsu
• Kawarimi: free local AI character swap for Windows and Linux, no cloud · jutsustudio.com/kawarimi
• EdgeTAM Core ML: EdgeTAM video object tracking in Core ML, on the Apple Neural Engine
• edgetam-onnx: the same tracker in Rust on ONNX Runtime CUDA, for Windows and Linux
• Higgsfield Gateway: a budget-capped Higgsfield API gateway on Cloudflare Workers · jutsustudio.com/blog/higgsfield-api-production
• Seedance 2.5 Prompt Skill: a Claude skill that writes Seedance 2.5 prompts for Higgsfield · jutsustudio.com/blog/seedance-prompt-skill
• AI video API prices: what one second of AI video costs at 30+ providers, updated daily · jutsustudio.com/video-api-pricing
• Open Source Higgsfield: every #OpenSourceHiggsfield challenge project, by category · jutsustudio.com/open-source-higgsfield
• awesome-higgsfield: tools, SDKs, agent skills and apps for Higgsfield

License

MIT © Content Factory contributors

_{Made for people who would rather tell the story than edit the timeline.}