One canvas, every medium
Image, video, audio, text, and 3D stay visible as typed steps—not buried in prompt history.
Wire the graph once, then change the inputs instead of rebuilding the process.

Open source · bring your own keys
Connect models, media, and repeatable steps on one canvas. Change the inputs, rerun the graph, and keep the useful mess that got you there.
Build the recipe, keep the experiments
Image, video, audio, text, and 3D stay visible as typed steps—not buried in prompt history.
Wire the graph once, then change the inputs instead of rebuilding the process.

Cross or zip lists into prompt constructors, price the matrix before it runs, and inspect the results together.
Every variation keeps the path that made it.

Bring Gemini, OpenAI, Anthropic, Replicate, fal, Kie, WaveSpeed, ElevenLabs, Ollama, or ComfyUI.
Share a workflow as a link, save it as a file, or run it headlessly.

Every workflow is a file that runs without the editor: from the CLI, over one HTTP call, or from any MCP client — the same engine and the same results, with long runs that outlive the request that started them.
# CLI
npm run workflow -- ./flow.json --input "Photo=@./shot.png"
# HTTP
curl -X POST localhost:3000/api/run \
-H 'Content-Type: application/json' -d @flow.json
# MCP
claude mcp add current -- node scripts/mcp-server.mjsThe full canvas, right here. Projects persist in this browser's storage and can be exported as .json at any time. Your provider keys stay in your browser and ride along only on your own requests — nothing is stored server-side.
Open the canvasThe same app from source, on your machine. Local adds what a hosted page cannot: