nodetool

NodeTool is an open-source creative workspace in which agents can construct, run, inspect, and repair multimodal workflows for image, video, audio, and text production. Users can start from a natural-language brief, review generated storyboards and stills, animate approved shots, add narration and music, edit on a timeline, and export MP4 video. The same project can be edited manually or through an agent. The studio provides several connected editing surfaces. The storyboard editor organizes shots, action, camera direction, motion, duration, narration, and style information, and supports revising individual shots without regenerating the whole sequence. The script and voice tools support character-specific provider, model, and voice selection, saved takes, word timings, and caption assembly. The timeline allows video and audio tracks to be arranged, trimmed, layered, and exported, while clips can remain linked to generating workflows. The sketch editor provides layered compositions with masks and blend modes, and the 3D editor can create scene references for image or video generation. Underlying these surfaces is a typed node-graph editor. Users can connect nodes, inspect intermediate outputs, swap models, and rerun only changed parts. Workflows can combine image, video, audio, and text operations such as masking, inpainting, outpainting, relighting, upscaling, layering, and compositing. Projects retain the graph, inputs, assets, and edits together. Model access is designed around user-owned credentials and local or provider-hosted models. The documentation lists providers including OpenAI, Anthropic, Gemini, FAL, KIE, Replicate, ElevenLabs, and HuggingFace, with local support through Ollama, MLX on Apple Silicon, and GGUF models. API calls are made directly to the selected provider using the user's key. Hardware requirements for local models depend on the model being used. Agent features include workflow construction, graph validation, custom mini-app generation, failure handling with retry or repair choices, cost budgets, and OpenTelemetry spans for LLM calls with token and cost information. The agent toolbelt is exposed through MCP, allowing external MCP clients such as Claude Desktop, Claude Code, Codex, or OpenCode to drive workflows and editing surfaces. A CLI can run workflows, serve the API, and provide interactive agent chat after a provider is configured. Developers can extend NodeTool with TypeScript or Python custom nodes, add provider integrations, write JavaScript nodes in a QuickJS sandbox, and use the built-in vector store for document search. Workflows can also be wrapped in simple mini-app interfaces for teammates. Deployment options include self-hosting with Docker and using remote GPU workers such as RunPod or Vast. The repository includes the TypeScript runtime and workflow kernel, providers and nodes, a React web UI, an Electron desktop shell, a React Native mobile app, documentation, and runnable workflow examples. NodeTool Studio is available for macOS, Windows, and Linux, while a hosted cloud offering is described as alpha. The project is licensed under AGPL-3.0.