creatorforge

The AI content team you own — not one you rent for $15K/month.

---

Business owner? Creator? Ask yourself:

• Are you paying (or about to pay) five figures a month for a "done-for-you AI content team"?
• Does that team's "proprietary system" run on someone else's servers, with your voice, audience, and data living there?
• When the retainer ends, do you keep anything — or does the system walk out the door with them?
• Do you actually want to film 2–4 hours a week and have everything else handled — without renting the engine that does it?

If you nodded, here's the deal: creatorforge is that content team, as open software you own and run yourself. No retainer. No 50-spots-only. No "proprietary" black box. It's free, Apache-2.0, and runs on your hardware and your model.

What you get

The same deliverables an agency installs — generated from one command:

• Voice profile — it learns your style from your past posts (no model training, nothing uploaded — just measurable style features).
• Niche research brief — content pillars, audience pains, keyword targets.
• Content ideas — a steady pipeline of format × angle ideas so you never face a blank page.
• Hooks — proven scroll-stopping formulas, written in your voice.
• Scripts — full, structured scripts sized to each platform.
• Captions — on-screen overlay text + timed SRT subtitles.
• Thumbnail concepts — headline / visual / emotion / layout, rendered as actual SVG mockups.
• Multi-platform packaging — one idea, tailored for YouTube, Shorts, TikTok, Reels, X, and LinkedIn (each platform's limits respected).
• Posting calendar — your ideas scheduled at your cadence.

You film. creatorforge does the rest.

How it's different (the honest version)

	The agency offer	creatorforge
Price	~$15K/month retainer	free, open source
Where it runs	their cloud	your hardware
Your voice/data	lives on their system	never leaves your machine
When it ends	you keep nothing	you own it forever
The "system"	proprietary black box	readable code you can audit
Results promised	"1 billion views"	a great engine — the views are on you

No inflated view-count promises here. creatorforge gives you a genuinely strong content system; what it can't do is guarantee virality, and anyone who guarantees that is selling something. It runs fully offline on a deterministic engine, and gets sharper the moment you point it at a local model (Ollama) or a cloud one — your call.

Quick start

pip install -e .

# 1. Learn your voice from a folder of your past posts/scripts
creatorforge profile ./my_posts/ --out voice.json

# 2. Run the whole team on a topic, across your platforms
creatorforge pipeline --topic "why you should own your AI stack" \
    --niche "AI for business" --platforms youtube,tiktok,x,linkedin \
    --voice voice.json --start 2026-07-06 --out plan.json

# Or one piece at a time
creatorforge hooks    --topic "cold email that converts" --voice voice.json
creatorforge script   --topic "cold email that converts" --platform youtube_shorts --voice voice.json
creatorforge thumbnail --topic "cold email that converts" --svg thumb     # writes thumb-1.svg …

# Use your own local model to sharpen the prose (nothing leaves your machine)
creatorforge script --topic "..." --provider ollama --model llama3

Run python demo.py to watch it learn a voice and generate a full multi-platform plan end to end.

Platforms

Platform	Aspect	Sweet spot	Tailored output
YouTube	16:9	~10 min	title + description + script
YouTube Shorts	9:16	≤60 s	short script + captions
TikTok	9:16	~27 s	caption + hashtags + script
Reels	9:16	~60 s	caption + hashtags + script
X	16:9	—	≤280-char post
LinkedIn	1:1	~90 s	long-form post

Local models — every modality, on your hardware

creatorforge drives the best open-source, local, free model your machine can run, in every modality — and degrades cleanly when you don't have the big one. Ask it what it can do:

creatorforge capabilities

Modality	What it does	Open model (GPU)	Runs with no GPU
Text	hooks, scripts, ideas in your voice	any Ollama model (auto-picks your best)	✅ via Ollama / template engine
Transcription	footage → text → repurposed posts	faster-whisper large-v3	✅ base/small on CPU
Voice	voiceover + voice cloning	XTTS-v2 (clone)	✅ Piper (CPU)
Image	photorealistic thumbnails	FLUX.1-schnell / SDXL-Turbo / SD 1.5	✅ real raster PNG (PIL) → SVG
Video	finished short, captions timed	LTX-Video / CogVideoX (text-to-video)	✅ assembled MP4 (ffmpeg) / animated GIF
Audio	voiceover + music, leveled	MusicGen beds	✅ ffmpeg mix / stdlib WAV

The recommended model in each row is chosen to fit your VRAM (recommend() ladders from a 24 GB GPU down to CPU). Photorealistic images, cloned voices, MP4s, and generated b-roll need the model/tool installed — but there's always a real, local fallback so nothing hard-fails: a composited raster thumbnail, an animated GIF, a synthesized WAV. Nothing is ever sent to a cloud API.

# transcribe footage with local Whisper
creatorforge transcribe talk.mp4

# generate a thumbnail (diffusion if you have it, else a real raster PNG)
creatorforge image --topic "owning your AI stack" --voice voice.json --out thumb

# produce a finished short from a script (MP4 with ffmpeg, else animated video)
creatorforge video --topic "owning your AI stack" --platform youtube_shorts --out short

# voiceover (Piper, or clone your voice with XTTS) and an audio track
creatorforge voiceover --from-script script.json --out vo.wav --speaker my_voice.wav
creatorforge audio --voiceover vo.wav --music bed.mp3 --out track

# the whole production in one command: plan + thumbnail + video + audio + outbox
creatorforge produce --topic "owning your AI stack" --niche "AI for business" \
    --provider ollama --out ./production/

Long-form & cinematic production (5–15 min)

Short clips are one mode. creatorforge also plans and assembles long-form video — documentaries, video essays, dev logs, promos — structured the way the industry actually does it. It fuses four things:

• a format (creatorforge formats) — the proven beat order for documentary / video-essay / devlog / promo
• a cinematic style (creatorforge styles) — pacing, shot vocabulary, color, and music mood drawn from the grammar of prestige docs, kinetic vlogs, trailer-cut blockbusters, slow-burn arthouse, true-crime, and more
• the algorithm playbook — what each platform actually rewards (watch time / AVD, completion rate, the first-30s hook, re-hooks every ~40s, chapters) turned into concrete production directives
• your voice (+ an optional local model to polish the narration)

…into a full plan: acts → scenes → timed shots → narration → chapters → a music & SFX cue sheet → titles and thumbnails, sized to your target runtime.

creatorforge formats        # documentary, video_essay, devlog, promotional
creatorforge styles         # epic_doc, true_crime, kinetic_vlog, blockbuster, arthouse_slowburn, …
creatorforge longform --topic "owning your AI stack" --format documentary --style epic_doc --minutes 12 --out plan.json
creatorforge studio   --topic "owning your AI stack" --format documentary --minutes 10 --provider ollama --out ./film/

Generative music and sound effects come from local open models (MusicGen / AudioGen) when your GPU can run them, with a synthesized bed as a fallback so a cut always has a track.

On "Netflix-level": creatorforge plans and directs at that structural level — story structure, shot lists, pacing, sound design, retention engineering — and assembles the cut with whatever render models you have. Cinema-grade footage and audio come from the heavy open models (FLUX / LTX-Video / MusicGen / XTTS) on a capable GPU, which the engine drives; on a laptop you still get a complete, correctly-structured cut with real generated assets. The structure is studio-grade everywhere; the render fidelity scales with your hardware.

Punch above your weight class: real photos, designed

You don't beat a giant cloud image model on a laptop by out-rendering it — you out-source and out-design it. creatorforge pulls real, no-watermark photography and composites it into pro visuals, which reads as more professional than weak CPU diffusion ever will.

• Your own library, offline. Index image folders you already have and search them by keyword — multiple related shots per scene, your images, zero watermarks, zero licensing questions. Optionally caption them with local llava for smarter matching.
• Free, no-key, no-watermark CC stock. Openverse / Wikimedia Commons, each result carrying its license + attribution so you stay compliant. Watermarked sources (Getty/Shutterstock previews) are never queried.
• Designed, not dumped. A sourced hero photo gets cover-cropped, color-graded, given a legibility scrim, and set with a bold headline — a real editor's thumbnail.

# build an offline, searchable index of your image folders (your own = no watermark)
creatorforge assets index ~/Pictures ~/b-roll --out assets.json

# pull multiple related, key shots for a scene
creatorforge assets search "datacenter server racks" --index assets.json -k 6
creatorforge assets search "founder at laptop" --index assets.json --online   # + free CC stock

# build a thumbnail composited over a real sourced photo
creatorforge image --topic "owning your AI stack" --assets assets.json --out thumb
# studio/produce auto-source a hero from your library:
creatorforge studio --topic "..." --format documentary --assets assets.json --out ./film/

Licensing, honestly: your own library is yours. Openverse/Wikimedia results are Creative Commons / public-domain and may require attribution — creatorforge surfaces the license and attribution string with every result so you can credit correctly. It never pulls watermarked or paid-preview imagery.

Direction & engagement craft (multi-cam, cool shots, retention)

Long-form plans don't just have scenes — they have coverage and engagement built in:

• Multi-camera shot generation. Every beat gets a shot list across A/B/C/detail/drone/POV cameras with motivated camera moves — dolly-in, push-in, crane, gimbal track, drone reveal, dolly-zoom (vertigo), orbit, speed-ramp — and the showy "cool shots" land where they matter (cold opens, turns, climaxes). The plan tells you exactly which cameras a production needs.
• Engagement craft, encoded. Each beat carries a retention move drawn from two lineages: great-filmmaker grammar (in-media-res, but/therefore causality, show-don't-tell, escalating stakes, tension-release, match cuts, the Kuleshov effect) and modern retention tactics (front-load the payoff, escalate every segment, reset the hook on a cadence, concrete stakes/numbers, no dead air, tease what's coming, emotional payoff). Structure isn't just correct — it's sticky.

It's all in the longform/studio plan: scenes[].shots (cams + moves), scenes[].retention_move, multicam, and engagement_plan.

Make content for your repos — and grow like the AI companies that made it

Point creatorforge at a repository and it writes the content for it; point it at an owner and it does the whole catalog.

# one repo -> multi-platform content (reads the README, derives what it is)
creatorforge repo ./codegraph-mcp --format promotional --out plan.json
creatorforge repo cognis-digital/agentledger --longform --format documentary   # via gh

# every repo of an owner, batched
creatorforge repos --owner cognis-digital --format promotional --out ./repo_content/

# a 30-day launch strategy that mirrors how the big AI companies actually grew
creatorforge growth ./codegraph-mcp

The growth playbook encodes the plays that repeatedly worked for the AI companies that broke out — a runnable demo over a pitch deck, building in public, a benchmark moment, developer-first distribution, founder-led content, free/open core, naming the category — and turns them into a concrete day-by-day launch calendar for your project. The edge was never the secret; it's executing these consistently, which is exactly what the engine makes cheap.

Wire it into your stack (MCP)

creatorforge ships an MCP server, so Claude, an internal orchestrator, or any MCP-capable agent can drive it directly:

creatorforge serve     # JSON-RPC over stdio

Tools: profile_voice, generate_ideas, write_hooks, write_script, thumbnail_concepts, package_for_platform, run_pipeline. Hand the packaged posts to your scheduler/poster of choice and the loop is closed.

Testing

pip install -e ".[dev]"
pytest -q          # 49 tests

License

Apache-2.0. © Cognis Digital. The whole engine is open — read it, fork it, run it on your own models. You own your content team.

Status: v0.1 — runnable and tested. Short-form + long-form (5–15 min) production; text, transcription, voice, image, video, music/SFX backends wired with hardware-aware selection and CPU fallbacks; format/style/algorithm intelligence baked in. Roadmap: ComfyUI image/video backend, true text-to-video assembly, A/B title scoring, scheduled auto-publish.