PYB-CLI - Minimal AI Agent

A minimal philosophy-based AI agent implementation that delivers complex problem-solving capabilities with the least amount of code.

Project Overview

PYB-CLI is an AI agent implementation that follows the minimalist agent philosophy, embodying the core concept of "letting LLMs do what they do best." This implementation achieves core functionality with minimal code, fully leveraging LLMs as general problem-solving engines.

Core Philosophy

1. "Let LLMs Do What They Do Best"

The core of the minimalist agent philosophy is to fully leverage LLMs as general problem-solving engines, reducing artificial complex architectures that limit LLM capabilities - everything is a scaffold for LLMs.

In traditional complex frameworks, we often design complex state management, execution flows, and architectural layers, but these artificially complex architectures may actually limit LLM's intelligent decision-making capabilities. Minimalist philosophy advocates for letting LLMs themselves become the "brain" for problem-solving, rather than relying on complex frameworks.

2. "Minimum Necessary Complexity" Principle

Only keep elements essential for problem-solving:

Essential elements: messages (conversation history), tools (capability extensions), LLM (decision engine) Non-essential elements: complex state management, fixed execution flows, complex persistence mechanisms

3. "Trust LLM Intelligence" Principle

Believe that LLMs can handle complex information integration and decision-making:

• No need to artificially segment state fields
• LLMs can extract required information from complete history
• Let LLMs decide how to organize and use information

4. "Transparency is Reliability" Principle

All decision processes are visible to humans:

• Conversation history can always be checked to understand decision-making
• Facilitates human supervision and intervention
• Easy to locate issues when problems occur

Design Philosophy

• Minimum Necessary Complexity - Only keep elements essential for problem-solving
• Trust LLM Intelligence - Fully leverage LLM's intelligent decision-making capabilities
• Transparency is Reliability - All decision processes are visible to humans

Core Features

• Minimal Implementation - Complete agent system with minimal files
• Dynamic Planning - Implicit state management based on message history
• Tool Extensions - Core capabilities for file operations, shell execution, etc.
• Feedback Loop - Intelligent cycle of perception → analysis → decision → execution → observation → replanning
• Easy to Understand - All states visible in message history for debugging and understanding
• Multi-Model Support - Supports multiple mainstream AI models, flexible for different scenarios

Supported Models

PYB-CLI supports multiple mainstream AI models that users can choose based on their needs:

Kimi Series:

• Kimi K2 Turbo (kimi-k2-turbo-preview) - Suitable for Chinese scenarios and domain-specific tasks
• Kimi K2 0711 Preview (kimi-k2-0711-preview) - Kimi model with vision capabilities

Claude Series:

• Claude 3.5 Sonnet (claude-3-5-sonnet-latest) - Suitable for complex reasoning and general tasks
• Claude 3.5 Haiku (claude-3-5-haiku-latest) - Lightweight and efficient model

GPT Series:

• GPT-4o (gpt-4o) - Balanced choice for cost-effectiveness and general tasks
• GPT-4o Mini (gpt-4o-mini) - Lightweight and efficient model
• O1 (o1) - Advanced model with reasoning capabilities

DeepSeek Series:

• DeepSeek Chat (deepseek-chat) - Suitable for coding tasks and cost-sensitive scenarios
• DeepSeek Coder (deepseek-coder) - Optimized specifically for code tasks

Qwen Series:

• Qwen Max (qwen-max) - Alibaba's Qwen high-performance model
• Qwen Plus (qwen-plus) - Alibaba's Qwen balanced model

GLM Series:

• GLM-4 (glm-4) - Zhipu AI's high-performance model

MiniMax Series:

• MiniMax Abab6.5s Chat (abab6.5s-chat) - MiniMax high-performance model

Llama Series (Ollama):

• Llama 3 (llama3) - Suitable for local deployment and privacy protection scenarios
• Llama 3.1 (llama3.1) - Updated version of Llama model

Model Selection Guide

• Complex Reasoning Tasks: Recommend Claude 3.5 Sonnet or O1
• Cost-Sensitive Scenarios: Recommend DeepSeek Chat, DeepSeek Coder, or Llama 3
• Chinese-Specific Tasks: Recommend Kimi K2 Turbo or Qwen Max
• Local Deployment Needs: Recommend Llama 3 or Llama 3.1 (Ollama)
• General Balanced Choice: Recommend GPT-4o or GLM-4
• High-Performance Needs: Recommend Kimi K2 0711 Preview or Qwen Max
• Lightweight Tasks: Recommend Claude 3.5 Haiku or GPT-4o Mini

Supported Tools

• bash - Execute shell commands
• read_file - Read file contents
• write_file - Write file contents
• edit_text - Edit file contents

Installation and Usage

As CLI Tool

# Run directly (no installation required)
npx pyb-ts

# Or install globally and use
npm install -g pyb-ts
pyb

Security Notice

Important: This tool requires you to configure your own API keys. Never share your API keys with others, and ensure they are kept secure.

First-Time Setup

After installation, you need to configure at least one AI model:

# Interactive configuration (recommended)
pyb add

# Or configure a specific model
pyb add gpt-4o --api-key YOUR_API_KEY --provider openai

# Set as default model
pyb set-default gpt-4o

# Set model pointers for different scenarios
pyb set-pointer main gpt-4o
pyb set-pointer task deepseek-chat
pyb set-pointer reasoning o1

Make sure to replace YOUR_API_KEY with your actual API key from the AI provider.

Getting Started

Set up your API key as an environment variable:

export ANTHROPIC_API_KEY=your-anthropic-key
# or
export OPENAI_API_KEY=your-openai-key
# or
export KIMI_API_KEY=your-kimi-key

Run the CLI:

pyb

Use configuration commands to manage models:

pyb add                     # Interactive mode
pyb add gpt-4o --api-key sk-xxx --provider openai
pyb list                    # List configured models
pyb set-default claude-3-5-sonnet-latest
pyb set-pointer main gpt-4o
pyb set-pointer task deepseek-chat
pyb switch                  # Switch between configured models

Note: Providers that currently support direct API key reading from environment variables include Anthropic, OpenAI, Kimi, DeepSeek, and Ollama. Other providers (such as Qwen, GLM, MiniMax) need to be configured through configuration files or interactive configuration wizard.

License

MIT