Presets
Save a subgraph as a reusable, parameterized preset and use built-in model templates.
Custom Nodes
Add new node behavior by dropping a Python file into custom_nodes/ — hot-reloadable, no frontend changes.
Subgraphs
Collapse a selection into a reusable block you can open, edit, and expand again — with every instance sharing one definition.
Plugins
Install plugin packs of educational nodes, and learn how to write and publish your own.
Device Backends
How CodefyUI selects and falls back across CPU, CUDA, MPS, and ROCm — plus the experimental native-MLX inference spike.
Plugin Frontend Extensions
Ship a JavaScript bundle with your plugin so it can add UI widgets, inspect graphs, and drive the editor — the foundation for Graph Copilot and similar tools.
API Reference
Backend routes and authentication requirements for nodes, graphs, runs, sweeps, published apps, plugins, optional packs, source control, files, media, and the LLM proxy, plus request limits, error formats, and the execution WebSocket protocol.
Graph Copilot
Use an AI assistant to generate and edit node graphs through the plugin frontend extension API and LLM streaming proxy.
Architecture
How CodefyUI is put together — backend-authoritative node definitions, WebSocket execution, topological scheduling, the frontend's stores and run events, and the registry/plugin system.
Execution Architecture
How a run gets from submission to a result when it can execute in the server process or on a worker agent — the control plane, executors, worker agents and run subprocesses, slots and placement, the event path, isolation, storage and security. The reasons, invariants and open questions are in Execution Design Notes.
Execution Design Notes
Why runs execute the way they do when a server can hand them to worker agents — the goals, the six invariants and the reason for each, the decisions with the alternatives they ruled out, the dated decision log, and the questions still open.
PythonScript Node
Write Python directly on the canvas — the contract, the Tier-0 import policy, the honest security model, and statistics recipes.
Training Memory
Fit a bigger run on one card — mixed precision, gradient accumulation, picking a specific GPU, what several GPUs change, what happens on an out-of-memory error, and the server's own memory budgets.