# CreatorOS Studio > Your own open-source Higgsfield, local: 430+ tools, presets, workflows - Page: https://www.whatwemade.app/p/creatoros-studio - What it is: Web app, for design, also music & audio - Made by: [thePloggy](https://www.whatwemade.app/u/theploggy) - Run it yourself: get the code and run your own copy - Live: https://theploggy.github.io/CreatorOS-Studio - Code: https://github.com/thePloggy/CreatorOS-Studio (Apache-2.0, 61 stars) - Languages: English - Status: in progress, 90% built - Wants contributors: Looking for testers with 16–24 GB GPUs, and feedback on the dance and sound studios - Tags: ai-influencer, ai-video, character-consistency, comfyui, allinonestudio - Upvotes: 3, comments: 3 ## About Built for the Comfy #OpenSourceHiggsfield challenge. Every model is open weights and runs locally: no API call, no credits, nothing leaves the machine. 50 functions across image, video, sound and editing ▎ 118 presets (dances, 3D dance studios, camera moves and effects, transitions, music styles, voices, reel formats) ▎25 built-in recipes, plus an internal app store: browse and import 262 official ComfyUI workflows (29 run without downloading anything), or import your own ▎ MCP server (15 tools) and agent skill, so Claude or any MCP client can drive the studio ▎MiniMax H3 ready: pre-wired, weights supplied by the user (its licence excludes the EU, UK, South Korea and the US) Tested on one NVIDIA GeForce RTX 5090 (32 GB) with ComfyUI 0.43 on Ubuntu 26.04. The GitHub README lists, for every function, the cold-run time, the median over real renders and the peak VRAM. ## How it was made - Claude Code, 70% of the work: code, maintenance, watch - Claude, 20% of the work: hardware setup and plan - ChatGPT, 10% of the work: original idea + first guidelines - Stack: React, Tailwind, TypeScript - MCP servers: GitHub ## What you need to run it - To use it: a good graphics card ## Starting brief Written by Claude Code for building your own version. INSTALLATION PROMPT: Install CreatorOS Studio on this machine from https://github.com/thePloggy/CreatorOS-Studio and get it running locally. 1. Check the machine first and report: OS, NVIDIA driver and GPU VRAM (nvidia-smi), Node 22.5+, git, ffmpeg, Python 3.12 (uv if present), free disk space. Stop and tell me what is missing before installing anything system-wide. 2. git clone https://github.com/thePloggy/CreatorOS-Studio creatoros-studio && cd creatoros-studio, then read README.md (Install, Setup wizard, Smaller GPUs, Configuration). 3. Run `npm run setup -- --check` and show me the report: ComfyUI location, missing custom nodes, missing models with their sizes, and the VRAM fit for each feature group. 4. Ask me which feature groups I want and recommend the ones that fit my GPU, using the exact group ids the setup report shows (e.g. photo, photo_klein, retouche, video_wan, video_ltx, motion, musique, parole). If my card has less than 24 GB, use the light profile (`--vram `). 5. Before downloading models, remind me that FLUX.2 Klein and LTX-2.5 come from gated Hugging Face repos: I must accept their licences on the model pages and export HF_TOKEN. Never print or store the token. 6. Run `npm run setup -- --groups ` and confirm with me before each step that writes or downloads (ComfyUI, custom node packs, models, Ollama models). Do not install MiniMax H3: its licence excludes the EU, the UK, South Korea and the US. 7. Restart ComfyUI so the new node packs load, check it answers on http://127.0.0.1:8188, and that Ollama answers on http://127.0.0.1:11434. 8. npm ci, npm run build, npm start, then open http://127.0.0.1:8787 and confirm the machine panel shows ComfyUI as connected. 9. Render one test photo from the UI or with `npm run studio -- photo "a woman reading by a sunny window" --wait`, and tell me where the file is. 10. Optional, only if I say yes: voice-over (kokoro venv), voice cloning (chatterbox venv), character training (ai-toolkit), and the MCP server (`claude mcp add creatoros-studio -- node /scripts/agent/mcp.mjs`). Finish with a short summary: what is installed, disk space used, which features are available on my GPU, and how to start the studio next time. --- FROM SCRATCH RECIPE: Build a web app called CreatorOS Studio: a local, open-source AI studio that turns one recurring character into photos, videos, dances, music and finished vertical reels, running entirely on the user's own GPU through a local ComfyUI. Stack: Node 22 + Fastify server (node:sqlite for jobs, assets and provenance), React + TypeScript + Tailwind front built with Vite, two layouts (desktop and phone) over the same API. The server never renders itself: it fills ComfyUI API-format workflows (JSON recipes, injected by node id), submits them to ComfyUI, waits, and stores every output with its recipe, seed, LoRAs and final prompt. Core features: - Characters: generate a reference sheet from a description, train an identity LoRA (FLUX.2 Klein / Krea 2) from the UI, then reuse that face everywhere. - Image: text to image, edit by instruction (Qwen Image Edit), inpaint, remove background, upscale, restore, relight, expand, multi-angle. - Video: image to video (Wan 2.2, LTX-2.5 with sound), text to video, camera moves, re-shoot an angle, motion control and character swap (Wan Animate, SCAIL-2), one-click LTX effects (refine details, day to night, unblur, remove compression). - Dance: Mixamo FBX played by a grey mannequin in one of 20 Blender studios (headless Blender script), the character placed on the first frame and relit, LTX-2.5 Layout-to-Render, then an identity swap. - Sound studio: music with lyrics (ACE-Step 1.5, tempo/key/vocals, two takes), text to speech, voice clone and changer, sound effects. - Editor: multi-clip timeline exported with ffmpeg, music ducked under voices, -14 LUFS master, generated LTX transitions with sound. - One-click reel: an LLM writes the shots, the code renders and edits them. - An in-app catalogue to import official ComfyUI workflows, a CLI and an MCP server (stdio, no dependency) so agents can drive the studio. Rules: - Open-weights models only, nothing leaves the machine. - One GPU, one render at a time, with a visible queue and time estimates. - A public demo mode: owner key for writes, per-visitor hourly limits, every prompt checked against a strict content filter (no minors, no sexual content) before and after any LLM rewrite, and visitor uploads only if no person is detected. - Every recipe and model listed in an installer manifest with its licence. Shared on WhatWeMade: https://www.whatwemade.app/p/creatoros-studio