Most LLM → image workflows stop at "send prompt, get image." This project treats an LLM as an active operator: the model authors workflows, edits nodes, chooses checkpoints, runs jobs, and inspects outputs — all against your ComfyUI instance. That shift turns ComfyUI from a passive renderer into an agent-driven creative loop that works with local or hosted LLMs and can run fully offline with local models.
What Sets It Apart
- Agent-native control plane: exposes ~150+ MCP tools that let an LLM manipulate ComfyUI node-by-node, validate workflows, and replay or rewind edits — so the model can iterate on graph structure instead of just sending prompts.
- Multi-provider parity: supports Claude, ChatGPT (Codex), Gemini, local LLMs (Ollama/LM Studio/llama.cpp), and OpenAI-compatible endpoints via a unified backend — so you can swap providers without losing panel features or tools.
- Local-first, offline-capable: auto-detects a local ComfyUI, manages models and custom node packs, and can run with fully offline local LLMs (no account required) — so sensitive or air-gapped setups remain usable.
- Installer packs & model expertise: ships curated installer packs, model-family generation guides, and a model registry so the agent auto-selects appropriate samplers, CFG, and checkpoints without trial-and-error.
- Two modes of integration: stdio MCP for local clients and a streamable-HTTP transport (with optional tunnel) for remote/hosted ComfyUI instances and panel-orchestrator setups.
Who it's for and tradeoffs
Great fit if you run or manage ComfyUI and want an LLM to perform higher-level workflows (authoring, debugging, batch sweeps, provenance tracking) without manual node fiddling. It benefits power users who need model management, reproducible workflows, or a sidebar agent that can visualize and modify the live canvas.
Look elsewhere if you only need a minimal relay (prompt→image) or you lack an LLM with tool-calling support — the system depends on tool-capable models for full automation. It also expects a running ComfyUI (local or reachable remote) and benefits from a GPU for timely generation. For tiny, single-command integrations a lightweight connector may be simpler.
