Pocket Flow: A 100-line minimalist LLM framework
- Expressive: Support (Multi-)Agents, Workflow, RAG, etc.—everything you’d expect from larger frameworks.
- Lightweight: Only the core graph abstraction in 100 lines. No dependencies or vendor lock-ins.
- Principled: Designed for clear modularity and separation of concerns.
- To install,
pip install pocketflowor just copy the source code (only 100 lines). - To learn more, check out the documentation. For an in-depth design dive, read the essay.
- 🎉 We now have a discord
For a new development paradigmn: Build LLM Apps by Chatting with LLM agents, Not Coding!
- 🧑 Human describe LLM App requirements in a design doc.
- 🤖 The agent (like Cursor AI) implements App your code automatically.
👈 How to set up Pocket Flow for LLM agents?
- For quick questions: Use the GPT assistant (note: it uses older models not ideal for coding).
- For one-time LLM task: Create a ChatGPT or Claude project; upload the docs to project knowledge.
- For LLM App development: Use Cursor AI.
- If you want to start a new project, check out the project template.
- If you already have a project, copy .cursorrules to your project root as Cursor Rules.
👈 How does Pocket Flow compare to other frameworks? Pocket Flow is purpose-built for LLM Agents:
- 🫠 LangChain-like frameworks overwhelm Cursor AI with complex abstractions, deprecated functions and irritating dependency issues.
- 😐 Without a framework, code is ad hoc—suitable only for immediate tasks, not modular or maintainable.
- 🥰 With Pocket Flow: (1) Minimal and expressive—easy for Cursor AI to pick up. (2) Nodes and Flows keep everything modular. (3) A Shared Store decouples your data structure from compute logic.
In short, the 100 lines ensures LLM Agents follows solid coding practices without sacrificing flexibility.
✨ Below are examples of LLM Apps built on top of Pocket Flow + Cursor AI:
| Formal App Name | Informal One-Liner | Difficulty | Learning Objectives |
|---|---|---|---|
| Youtube Summarizer | Explain YouTube Videos to you like you're 5 | ★☆☆ Beginner | Map Reduce |
| YC Adice Retriever | AI Paul Graham, in case you don't get in | ★☆☆ Beginner | RAG |
| Cold Opener Generator | Instant icebreakers that turn cold leads hot | ★☆☆ Beginner | Map Reduce |
- Do you want to create your own Python project? Start with this template
The 100 lines capture what we believe to be the core abstraction of LLM frameworks:
- Computation: A graph that breaks down tasks into nodes, with branching, looping, and nesting.
- Communication: A shared store that all nodes can read and write to.
From there, it’s easy to implement popular design patterns like (Multi-)Agents, Workflow, RAG, etc.
