nanoClaw

A tiny OpenClaw for educational purposes. A minimal AI agent that actually does things—runs shell commands, remembers across sessions—so you can see how the pieces fit together.

What it does

• Chat with Claude in a colorful CLI with configurable model (defaults to Haiku)
• Execute bash commands when the model decides it needs to (e.g. ls, grep, cat)
• Persistent memory stored in MEMORY.md—the agent remembers facts across sessions
• Extended Thinking so you can see the model's reasoning before actions
• Rich console output with colors, emojis, and formatted display

Quick start

# Create a virtual environment and install dependencies
python -m venv .venv
source .venv/bin/activate   # or `.venv\Scripts\activate` on Windows
pip install -r requirements.txt

# Add your Anthropic API key to a .env file
echo "ANTHROPIC_API_KEY=your-key-here" > .env

# Run
python main.py

Type your message, press Enter. Type exit or Ctrl+C to quit.

Project structure

File	Purpose
main.py	Main loop: chat, tool execution, message history
engine.py	Claude API wrapper with Extended Thinking
tools.py	execute_bash tool schema and implementation
memory.py	save_to_memory tool and markdown persistence
config.py	Centralized configuration (model, timeouts, etc.)
MEMORY.md	Stored memories in human-readable markdown
requirements.txt	Python dependencies (anthropic, rich, dotenv)

How it works

1. Configuration is loaded from config.py (model, timeouts, etc.)
2. On startup, the agent loads memories from MEMORY.md into the first message
3. You send a message; Claude may think, use tools, or reply
4. If it calls execute_bash, the command runs and the output is fed back
5. If it calls save_to_memory, the fact is appended to MEMORY.md
6. The loop continues until Claude responds with text and no tool use

Configuration

Edit config.py to customize:

• Model: Change from Haiku to Sonnet or Opus
• Token limits: Adjust max_tokens and thinking_budget
• Timeouts: Set bash command timeout
• Memory file: Change storage location
• Exit commands: Customize quit commands
• Log level: Set to "DEBUG" to see raw LLM responses, token usage, and detailed execution flow

Related

• OpenClaw — full-featured autonomous agent (multi-platform, browser automation, skills, etc.)
• nanoClaw is a stripped-down version to learn the core loop: LLM + tools + memory.