ComfyClaw: Noodle your way to custom agent workflows.
I’ve been building something called ComfyClaw, an experimental agent harness inside ComfyUI. It includes a full custom node suite, a CLI tool suite, an agent file system with prompts, and a local chat interface.
The idea is not to make a cutting-edge autonomous agent. It’s more of a visual workbench for exploring how agentic systems work under the hood. Instead of hiding everything inside code, ComfyClaw lets you see the loop: prompt construction, heartbeat checks, internet search, browser use, tool calls, short-term memory, long-term memory, safety checks, and a kind of subconscious planning step that suggests useful tools before the main agent acts.
It runs locally and is designed to be completely free to use. It can browse the internet, search the internet, and use local tools without needing paid APIs or API keys to protect. You can run it with local models through Ollama, and the tool system is just local Python scripts.
What interests me most is the memory side. Older interactions can be compressed into short-term memory, then eventually turned into long-term memory with importance scores, keyword associations, and relationship notes. So the agent is not just keeping a chat log; it starts to build a rough sense of its past and what mattered.
It’s still early and not especially fast, because ComfyUI is being used as a prototyping environment rather than a production runtime. LLMs can also still be unreliable with exact tool-call formatting, which is one of the things this project helps expose and test.
But that’s the point: ComfyClaw is for people who want to experiment with agent concepts, memory, tool use, and workflow design without needing to build a full agent stack from scratch.
GitHub: https://github.com/Steven-Hammon/ComfyClaw
#ComfyUI #AIAgents #AgenticAI #LocalAI #Ollama #OpenSourceAI #AIWorkflow llm