Open Generative Studio

An open-source, self-hostable AI creation studio. Generate images, video, lip-sync, cinematic clips and marketing/UGC ads from one canvas - with a node-based workflow builder, a per-user credit wallet, and a hosted MCP server so AI agents (Claude, ChatGPT, Cursor) can generate media through your instance.

Model inference is powered by Atlas Cloud - one API key gives your instance access to frontier image, video, audio and LLM models (FLUX, Qwen-Image, Kling, Seedream, DeepSeek and more).

Features

• Image Studio - text-to-image and image-to-image with reference images, multi-model picker, and a live per-generation cost pill.

• Video Studio - text-to-video, image-to-video and reference-to-video with multi-image / video / audio reference support.

• Lip Sync - portrait + audio (uploaded or recorded in-browser) to talking video.

• Marketing Studio - templated UGC / product-ad video generation.

• Flow - a node-based workflow builder (React Flow): chain LLM, image, video and audio nodes, fan out one prompt into N variations, and start from a library of pre-built community templates with baked example media.

• MCP server - a hosted Model Context Protocol endpoint with OAuth 2.1 one-click connect, so agents can call generate_image, generate_video, get_balance and more against your instance.

• Credit wallet - atomic Postgres credit ops (deduct / refund / true-up), append-only transaction ledger, optional Razorpay top-ups, promo codes, and a two-tier affiliate program.

• Production-hardened - refund-on-failure generation pipeline, async submit-and-reconcile for long video jobs, result re-hosting into Supabase Storage, rate limiting, and an admin business-metrics dashboard.

Related MCP server: Agnes MCP Server

Architecture (60 seconds)

Browser (studio canvas)
   |  POST /api/generate { model, prompt, ... }
   v
Next.js API route - Supabase auth -> atomic credit deduction -> dispatch
   |
   |  model_costs.source drives the dispatcher:
   |    'atlascloud' -> lib/atlascloud/client.js  (primary provider)
   |    'kie.ai' / 'muapi.ai' -> optional secondary adapters
   v
Atlas Cloud  (async prediction -> polled by the reconciler -> result re-hosted
              into Supabase Storage -> row completed / auto-refunded on failure)

• Frontend: Next.js 15 (App Router) + React + Tailwind.

• Backend: Next.js API routes + Supabase (Postgres, Auth, Storage).

• Inference: Atlas Cloud - the app talks to https://api.atlascloud.ai server-side only; API keys never reach the browser.

Quickstart

git clone https://github.com/nextgentrainingacademy88-max/open-generative-studio.git
cd open-generative-studio
cp .env.local.example .env.local   # paste your Supabase + Atlas Cloud keys
npm install
npm run build:packages
npm run dev                        # http://localhost:3000

Full walkthrough (Supabase migrations, storage bucket, model registration): SETUP.md.

Powered by Atlas Cloud

This project uses Atlas Cloud as its model inference provider:

• LLMs via the OpenAI-compatible POST /v1/chat/completions endpoint.

• Images via POST /api/v1/model/generateImage (FLUX, Qwen-Image, Seedream, ...).

• Video via POST /api/v1/model/generateVideo (Kling, Vidu, ...).

• Async predictions are polled at GET /api/v1/model/prediction/{id} and finalized by the built-in reconciler.

Set ATLASCLOUD_API_KEY in your environment and register the models you want in the model_costs table (source = 'atlascloud') - see supabase/migrations/0052_atlascloud_provider.sql for a commented example. Keys are read server-side only and are never committed to this repository.

Repository tour

app/                    Next.js App Router - marketing pages, studio, wallet,
                        affiliate portal, blog, /api routes (generate, workflow,
                        razorpay, mcp, cron)
components/             App shell, marketing sections, admin charts, UI primitives
lib/                    Server-side core: credit ops, pricing, dispatch
  atlascloud/client.js  Atlas Cloud adapter (images, video, chat)
  generate-core.js      Auth -> deduct -> dispatch -> refund-on-failure pipeline
  generation-reconcile.js  Async poll + re-host + true-up
  flow/                 Workflow engine (poll-driven, fan-out, CAS-claimed ticks)
  mcp/                  MCP tools + auth (API keys + OAuth via Supabase)
packages/               Vendored studio UI + workflow-builder (React Flow)
supabase/migrations/    Full schema: wallet, generations, model catalog, promo,
                        affiliate program, security hardening
scripts/                Template seeders + maintenance scripts
skills/                 Agent skill bundle for the MCP server

Credits & lineage

Started from Anil-matcha/Open-Generative-AI (MIT) and substantially rebuilt: credit-wallet SaaS backend, server-side provider dispatch, Flow workflow engine + builder UX, MCP server with OAuth, affiliate program, admin analytics, security hardening, and the marketing site.

License

MIT - free to use, self-host, modify and redistribute.