Features

• Small: nano_manus is about 1000 LOC.
• Using MCP: nano_manus supports loading MCP from docker, npx and Smithery
• Plan-then-Execute: nano_manus will gather your agents, make the plans and then assign the correct jobs to your agents

Use Cases

https://github.com/user-attachments/assets/33b90afc-2e30-4ab0-8988-bd7d6da065fb

• Find all .py file and explain them to me
• Give me the latest weather in SF in last 7 days and save it to csv

Welcome to give more use cases!

QuickStart

Setup

• Install NPM
• Install docker

Env

• Brave Search API key
• Jina API key
• OpenAI API key

Your .env should look like:

BRAVE_API_KEY=BSAxxxx
JINA_API_KEY=jina_xxxx
OPENAI_API_KEY=sk-proj-XXXXX

Run default nano-manus

uv sync
uv run examples/basic_planner.py

Abilities

• Search Web (mcp/brave-search, jina-ai-mcp-server)
• Local files operations (@wonderwhy-er/desktop-commander)
• Execute codes and commands in your computer (@wonderwhy-er/desktop-commander)
• (coming soon) Implementing CodeAct
• (coming soon) Read .pdf, .doc
• (coming soon) browser use
• (coming soon) multi-model router (claude, qwen, deepseek...)

Known Issues

• nano-manus is extremely unstable! My guess is gpt-4o is not that good at tool use.
• Unable to exit: seem like some MCPs will cause the problems of unable to exit the program when all the tasks were done.
• nano-manus will operate files and run command on the current dir of your local computer, make sure you don't run it on some important folders.