PocketFlow is a 100-line, zero-dependency Python framework for building LLM agents, multi-agent systems, workflows, and RAG pipelines, aimed at developers who want graph-based orchestration without the weight of larger frameworks.

What it is

PocketFlow lives in the Python LLM development ecosystem and reduces the whole framework down to a single file, pocketflow/__init__.py, which is 100 lines long. Its core abstraction is the Graph: nodes and edges, from which the documentation shows how to build popular design patterns such as Agents, Multi-Agents, Workflow, and RAG. The project ships under the MIT licence, carries 11,177 stars and 1,213 forks, and is published as a Python package named pocketflow.

The concrete problem it solves is the bloat and lock-in of incumbent LLM frameworks. The project's own comparison table measures LangChain at 405K lines and +166MB, CrewAI at 18K lines and +173MB, SmolAgent at 8K lines and +198MB, LangGraph at 37K lines and +51MB, and AutoGen at 7K lines core-only and +26MB, against PocketFlow's 100 lines and +56KB. PocketFlow replaces those installations with zero dependencies, zero app-specific wrappers, and zero vendor-specific wrappers, so the application talks to whatever model or vector store it chooses without a framework-owned adapter layer.

Key capabilities

• Core Graph abstraction of nodes and edges, from which the documented patterns for Agents, Multi-Agents, Workflow, and RAG are built.
• The entire framework is 100 lines in pocketflow/__init__.py, small enough to read and copy rather than install.
• Zero dependencies and zero vendor lock-in, with no bundled OpenAI, Pinecone, or similar wrappers in the comparison table.
• Installable as the pocketflow PyPI package, or vendored by copying the source file directly.
• Agentic coding workflow, in which AI coding agents such as Cursor AI generate agent code against the framework, described in the README as a 10x productivity boost.
• Ports to TypeScript, Java, C++, Go, Rust, and PHP, so the same graph model carries across language ecosystems.
• Documentation in Chinese, Spanish, Japanese, German, Russian, Portuguese, French, and Korean translations of the cookbook README, alongside the English original.

Who uses it and how

• Developers practising agentic coding, where an AI coding agent such as Cursor AI writes the agent implementation against PocketFlow's small surface.
• Teams that need LLM features but refuse framework-owned vendor abstractions, and therefore wire their own model and storage clients into graph nodes.
• Polyglot engineering groups that start in Python and reuse the same design patterns through the TypeScript, Java, C++, Go, Rust, or PHP ports.
• International developers, served by the nine cookbook translations, who need the pattern documentation in their own language.
• Community members who join the project's Discord to compare implementations with other developers building on PocketFlow.

Getting started

Install with pip install pocketflow, or copy the 100-line source from pocketflow/__init__.py into the project. The project homepage at https://the-pocket.github.io/PocketFlow/ and a video tutorial carry the design-pattern documentation.

How it compares

The README positions PocketFlow directly against LangChain, CrewAI, SmolAgent, LangGraph, and AutoGen, and the axes it uses are line count, installed size, and the number of abstraction layers, app-specific wrappers, and vendor-specific wrappers each one carries. PocketFlow claims the smallest footprint of the group at 100 lines and +56KB, with no app-specific or vendor-specific wrappers at all. It concedes scope rather than capability: the larger frameworks bundle tools and integrations that PocketFlow expects the developer to supply.

When to use it — and when not to

Choose PocketFlow when the goal is a small, readable, dependency-free graph layer and the team is willing to write its own model, storage, and tool integrations on top. Avoid it when the project needs prebuilt wrappers, batteries-included tools, or an opinionated application layer, because PocketFlow deliberately ships none. The trade-offs worth knowing are that the repository carries 75 open issues, that the README points outward to the documentation site, cookbook, and video tutorial for substance, and that a framework this thin leaves most integration and error-handling decisions to the application.