🧩 PocketFlow：10.6k Stars 的 100 行 LLM 框架，让 Agent 自己构建 Agent / PocketFlow: The 100-Line LLM Framework That Lets Agents Build Agents

Repo: The-Pocket/PocketFlow | ⭐ 10,635 Stars | 🛠 Python | 📦 56KB | MIT

你在看 LangChain 那 405K 行代码头大吗？CrewAI 一个 pip install 下去 173MB 就没了？PocketFlow 告诉你——100 行就够了。

不是"核心模块 100 行"，是整个框架就 100 行。去它的 __init__.py 看一眼，真的就 100 行 Python。零依赖、零厂商锁定、零臃肿。装个包 56KB，你手机空间都比它大。

对比一下：

装一下看看

pip install pocketflow

或者直接把源码复制到项目里——真的就一个文件。

真实用法：10 行搭一个多 Agent 工作流

from pocketflow import Node, Flow

class ResearchNode(Node):
    def exec(self, prep_res):
        # 调你的 LLM 做研究
        return "研究结果"

class WriteNode(Node):
    def exec(self, prep_res):
        # 根据研究结果写文章  
        return "文章草稿"

# 串起来
research = ResearchNode()
write = WriteNode()
research >> write  # 和 bash 的 pipe 一样

flow = Flow(start=research)
flow.run(shared_data)

对，就这么简单。>> 就是连接两个节点。

PocketFlow 的设计哲学是 Graph——所有 LLM 框架的核心抽象就是一个图。有了图，什么 Agent、Multi-Agent、RAG、Workflow、Supervisor、Debate 这些设计模式，全都能搭出来。项目里自带 30+ 教程，从 Chat 到 MCP 到 A2A 到 Voice Chat 全有：

# 拉个例子直接跑
git clone https://github.com/The-Pocket/PocketFlow.git
cd PocketFlow/cookbook/pocketflow-agent
python agent.py

谁该用？

• 被 LangChain 抽象恶心到的朋友
• 想搞懂 Agent 框架到底在干什么的新手（100 行代码看一遍就全懂了）
• 需要轻量级嵌入到现有项目里的场景
• 想用 Cursor/Claude Code 让 AI 帮你写 Agent 的（README 里叫这"Agentic Coding"）

English Version

🧩 PocketFlow: 10.6k Stars — 100-Line LLM Framework, Zero Bloat

Tired of LangChain's 405K lines and 166MB installs? PocketFlow is literally 100 lines of Python — zero dependencies, zero vendor lock-in, zero bloat. The entire framework fits in a 56KB package.

Install

pip install pocketflow

Or just copy the single file into your project.

Quick Start: 3-Node Research Workflow

from pocketflow import Node, Flow

class ResearchNode(Node):
    def exec(self, prep_res):
        return "research findings"

class WriteNode(Node):
    def exec(self, prep_res):
        return "article draft"

# Chain them together — the `>>` operator works like a pipe
research = ResearchNode()
write = WriteNode()
research >> write

flow = Flow(start=research)
flow.run(shared_data)

PocketFlow's core insight: every LLM framework is just a Graph. From that single abstraction, you can build Agents, Multi-Agent systems, RAG pipelines, Supervisor patterns, MCP tools, and more. The repo ships with 30+ cookbook examples covering all of them.

Who's it for?

• Developers who want to understand what their Agent framework is doing
• Anyone shipping lightweight LLM features into existing apps
• People who want their AI coding agent to build other agents ("Agentic Coding")

📡 Auto-generated project recommendation — not sponsored.