Overview
Noydflow is an agentic AI workflow platform where creators chain generative models on a visual canvas to produce images, video, music, voice and 3D assets. Instead of jumping between a dozen AI tools, users wire typed nodes together, from a reference image to a finished film with a soundtrack, and run the whole pipeline in one click. The same studio is exposed as a Model Context Protocol (MCP) server, so ChatGPT and other AI agents can call Noydflow's tools directly from a conversation.
The screens above are the product's design concept for the studio, the connected workflow canvas, the node editor and the output library.
What Makes It Agentic
• LLM Router node : A reasoning model sits inside the workflow, reads upstream inputs and writes the prompts and creative direction for downstream nodes. Users describe the intent once and the Router adapts it per model.
• Autonomous multi-step execution : Workflows run node by node across text, image, video and audio, with each output flowing into the next tool through typed connections.
• MCP server with OAuth : AI apps such as ChatGPT authenticate against Noydflow (RFC 9207 issuer identification) and gain access to the model catalog, uploads, previews and a Router "ask model" tool, turning the studio into a toolset other agents can operate.
• Server-side charging after results : Credits are only deducted once a provider returns a result, with per-model pricing and plan gating handled centrally rather than in the client.
Key Features
• 62 AI model nodes : Nano Banana, Seedance 2.5, Kling, Flux, Midjourney, Ideogram, Recraft, GPT Image, Grok Imagine, HeyGen avatars, ElevenLabs and Grok voices, Suno and Lyria music, Meshy 3D, Topaz and SeedVR upscaling, SAM masking, lip-sync, subtitles, depth maps and more.
• Node-based workflow builder : Built on React Flow with collapsed and expanded node states, visible connectable ports, and canvas checkpoints saved to the user's account.
• AI Clipping and Web Scrape : Turn long YouTube videos into short clips and pull structured data from websites, both gated by subscription plan.
• Creative library and history : Every output is stored with its prompt, model, workflow and format so it can be inspected and reused in the next pipeline.
• Bring-your-own API keys : Users can attach their own fal.ai keys, which are encrypted at rest and resolved server-side per request.
• Billing and plans : Stripe checkout, webhooks and fulfillment power credit packs and Starter, Plus and Pro subscriptions, with an admin console for credit sync, provider incidents and low-balance guards.
Technologies Used
• Frontend : React, Vite, TypeScript, React Flow, Tailwind CSS, shadcn/ui and Motion
• Backend : Node.js and Express with Zod validation and JWT (jose) auth
• Agent layer : Google Gemini for the Router node, the Model Context Protocol SDK and MCP Apps for the ChatGPT integration
• AI providers : fal.ai client plus direct integrations for video, voice, music and 3D models
• Data and auth : Supabase for Postgres, authentication, OTP verification and storage
• Payments and hosting : Stripe for billing, Railway for deployment, Resend for transactional email
Challenges and Learnings
• Orchestrating 60+ providers : Each model has different inputs, durations and failure modes. A shared provider layer with incident tracking and clear user-facing errors kept the canvas reliable as the catalog grew.
• Making a studio agent-callable : Exposing the product over MCP meant designing OAuth, scoped tool metadata, upload widgets and image previews that work inside a chat window, not just a browser.
• Fair, accurate billing : Charging only for results that came back, syncing admin credits and guarding against double-stacking required moving every charge decision to the server.
• Canvas usability at scale : Keeping nodes compact while leaving every port connectable led to a documented collapse pattern that all 62 nodes follow.
Outcome
Noydflow is live at noydflow.com . Noydflow turns a sprawling set of generative models into one coherent, agent-driven studio. Creators build reusable image-to-video-to-sound pipelines without code, and AI assistants can run the same tools through MCP, making Noydflow both a product for humans and a toolset for agents.
View project