Most LLM frameworks grow by bolting on a wrapper for every model, vendor, and use case until they balloon past 100K lines. PocketFlow bets on the opposite: the entire engine is roughly 100 lines expressing a single idea -- a graph of Nodes wired together by Flows -- and everything richer (agents, RAG, multi-agent systems) is a pattern you compose on top rather than a built-in module.
What Sets It Apart
- One abstraction, not a feature catalog. Because there is only the graph, the whole framework is learnable in an afternoon; there are no provider classes or chain types to memorize as the ecosystem churns.
- Zero dependencies, zero vendor lock-in. Nothing is pinned to a specific LLM SDK, so swapping providers or running fully offline is a change in your own node, not a framework migration.
- Small enough to be read by an AI. The core fits in a single context window, which is the point of its "Agentic Coding" workflow: a human designs the flow and a coding agent fills in the nodes.
- Size as evidence. Against LangChain (~405K lines) and CrewAI (~18K lines), the 100-line core is a deliberate argument that most framework bulk is incidental, not essential.
Where It Fits
In a field of heavyweight frameworks -- LangChain (~405K lines), LlamaIndex, CrewAI (~18K lines) -- PocketFlow occupies the minimalist extreme. It is closest in spirit to writing your own orchestration, but with just enough structure (the node-and-flow graph) to keep branching, retries, and multi-agent handoffs organized.
Great Fit / Look Elsewhere
Great fit if you want to understand exactly what your agent does line by line, value a dependency-free core, or are pairing with a coding agent that generates the node implementations for you. Look elsewhere if you need an out-of-the-box toolkit -- managed connectors, retrievers, observability, eval harnesses -- since PocketFlow deliberately ships primitives, not batteries, and leaves the LLM calls, parsing, and integrations for you to write.
