Start Anything

ComfyUI doesn't have a native "loop" construct - every queue is a fresh graph evaluation. Packs that want repeated, stateful execution (an agent that keeps chatting, a process that keeps refining something over several passes) have to fake it, usually with a node that remembers a value under a key and can either take a fresh input or replay what it remembered last time. start_anything is LLM Party's version of that primitive, and it's the generic one - it works with any data type at all.

Reading the shape of its inputs (a wildcard value, a key, and an is_reload switch), the pattern is: on a normal run, start_anything just passes init_any straight through as any. When is_reload is on, it instead reloads whatever was stored under key from a previous run, letting a workflow pick up where an earlier queue execution left off instead of starting cold every time. That's the building block you need for anything that has to survive across separate ComfyUI executions - a multi-step agent process, an iterative refinement loop, anything where "run it again" should mean "continue," not "restart."

Inputs and outputs that matter

• init_any - the wildcard (*) starting value. Connect literally anything: a string, a number, a whole workflow's worth of accumulated state.
• key - a string naming the slot this value is remembered under. This is what lets more than one loop coexist in the same workflow without stepping on each other.
• is_reload - flip this to resume from the stored value instead of taking a fresh init_any.

The output, any, is also wildcard-typed, so it plugs into whatever your loop's body actually consumes.

Installing it

Search comfyui_LLM_party in ComfyUI Manager and install, or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/heshengtao/comfyui_LLM_party

pip install -r requirements.txt inside your ComfyUI environment, restart. start_anything itself has no model or API dependency - it's pure workflow plumbing - but it ships as part of this large all-in-one pack, so you're pulling in everything else along with it. If you only want the lighter footprint, the README's only_api branch skips the local-model-loading dependencies.

Common issues

Because this node is a wildcard passthrough, ComfyUI's type checker won't catch a mismatch for you - if init_any is a string on one run and you accidentally wire in a completely different type on the next, whatever consumes any downstream will be the thing that errors, not this node. Keep the type consistent per key.

The other thing worth internalizing: key is the whole mechanism. If two separate loops in one workflow (or two copies of the same subgraph) reuse the same key string, they'll read and write the same stored slot and quietly interfere with each other - give each loop's start_anything its own distinct key, especially once a workflow grows past a single agent. And remember is_reload only does something useful if there's actually a prior stored value to reload - on the very first run of a fresh graph, there's nothing to resume from yet, so it behaves the same as a normal pass-through.