Reflow Studio

Your own AI image and video studio. Bring your own keys.

A self-hosted web app, REST API and MCP server that generates images and videos with 45+ models (Nano Banana, GPT Image, FLUX.2, Seedream, Kling, Veo, Seedance, Hailuo and more) through fal.ai, Kie.ai and the Higgsfield API. You pay the providers directly at their API prices. No subscription, no credits, no middleman.

Setup guide · MCP tools · Add a model · Architecture

Made with Reflow Studio

Every image below was generated with Reflow Studio for less than half a cent each.

Prompts and models: docs/assets/showcase/manifest.json.

Related MCP server: imagegen

Why

Hosted AI studios charge a monthly subscription and sell you credits on top of the same models you can call yourself. Reflow Studio gives you the same workflow on your own free-tier infrastructure:

Features

• Create images and videos with text, reference images, start and end frames, and model-specific settings. Live cost estimate before you click Generate.

• Cheapest provider routing. One model, several providers: the router picks the cheapest configured one and falls back when a provider is down.

• Library and assets. Every output is copied into your own storage and stays searchable by prompt and model.

• MCP server built in. Let your AI agent generate media: models_explore, generate_image, generate_video, jobs_wait, media_import_url and more.

• REST API with API keys and scopes.

• Keys stay yours. Provider keys are encrypted (AES-256-GCM) in your own database and never reach the browser or the MCP channel.

• Budget guard. Set a monthly cap in USD; Reflow Studio refuses jobs that would cross it.

• Owner-only by default. Only the e-mail addresses you list can sign in, enforced in the app and in the database.

Quick start

Option 1: Deploy to Vercel (about 5 minutes)

1. Click Deploy with Vercel above. Vercel copies the repo to your GitHub and creates a free Supabase database for you (pick region eu-west-1 if you are in Europe).

2. Fill in three values:

   • OWNER_EMAILS: the e-mail you will log in with.

   • REFLOW_SECRET: a random string of 32+ characters. Generate one with node -e "console.log(require('crypto').randomBytes(32).toString('base64'))".

   • MONTHLY_BUDGET_USD: your spend cap, for example 10.

3. When the deploy is done, open your site and go to /setup to create your account.

4. Paste a provider key in Settings → Providers (fal.ai, Kie.ai or Higgsfield API).

5. Generate your first image.

The database tables are created automatically during the build. Full guide, including "bring your own Supabase": docs/SETUP.md.

Option 2: Run it locally

git clone https://github.com/reflow-automations/reflow-studio-oss.git
cd reflow-studio-oss
corepack enable
pnpm install
cp apps/web/.env.example apps/web/.env.local
pnpm dev

Set ENABLE_MOCK_PROVIDER=true in apps/web/.env.local to try the full flow without any provider key or cost.

Connect your AI agent (MCP)

Create an API key in Settings → API keys, then:

claude mcp add --transport http reflow-studio https://<your-app>/api/mcp --header "Authorization: Bearer rfl_..."

Codex, Cursor, Claude Desktop and other clients: docs/MCP.md.

What it costs

Tech stack

pnpm monorepo: packages/core (model catalog as data, provider adapters, router, cost estimation) and apps/web (Next.js 16, Supabase, MCP server). Details in docs/ARCHITECTURE.md.

Contributing

Adding a model is one data entry in packages/core/src/catalog/models/. See CONTRIBUTING.md. Found a security issue? See SECURITY.md.

License and disclaimer

MIT. Built by Reflow Automations.

Reflow Studio is not affiliated with or endorsed by Higgsfield AI, fal.ai, Kie.ai or any model vendor. Product names are trademarks of their owners and are used only to describe compatibility. The MCP tool names follow Higgsfield's public MCP vocabulary so existing prompts and agent skills keep working.