The Problem
Chaining together multiple AI APIs for images, video, and text usually means a mess of Python scripts, JSON glue, and breakable shell commands. Managing API keys, prompt construction, and juggling outputs across providers is tedious and error-prone—especially if you want anyone else to use or understand your pipeline.
What This Does
node-banana gives you a visual node-based editor (check out src/components/WorkflowCanvas.tsx) to build generative AI workflows without wanting to throw your laptop out the window. Drag nodes for image generation, LLM prompt construction, or video synthesis onto an infinite canvas, wire them up, and execute the pipeline. The backend (src/app/api/generate/route.ts and friends) handles calling Google Gemini, Replicate, fal.ai, and others, all in dependency order.
Everything is local and portable. You can export and import workflows as JSON, and API keys go in .env.local (or just use the project settings). There’s a pile of test coverage in src/app/api/generate/tests and src/components/tests—so it’s not just a pretty UI demo.
Real-World Use
Say you want to generate a product shot with a swapped background, then annotate it for your marketing team. Drop a GenerateImageNode for Gemini, wire its output into a BackgroundSwapNode, and finish with an AnnotationNode. Save the workflow as JSON, or share it with someone who refuses to code. There are example images and templates in /examples/sample-images/ and /public/template-thumbnails/. Hell, you can even build the workflow from a text description if you feel lazy.
The Bottom Line
If you’re cobbling together AI tools and sick of duct tape, this is worth a look. Great for prototyping and demoing ideas without writing glue code. It’s not for running production pipelines at scale, but it absolutely beats a pile of brittle scripts for most use cases. If you like to see what’s happening, not just hope it works, try it.