Mini Manus
AI Coding Agent with Sandbox Execution
An AI agent that writes, executes, and iterates on code in isolated Docker containers. Describe what you want, get working code with live preview.
Overview
Mini Manus is a coding agent that goes beyond generation. It writes code, executes it in a sandboxed Docker container, observes the results, and iterates until the task is complete.
Challenge
Most AI coding assistants stop at generation. They produce code that might work, but verification is left to the user. Real programming is iterative: write, run, observe, fix. I wanted an agent that could close this loop autonomously.
Approach
Built a multi-step agent loop with AI SDK v5 and Claude Sonnet. The agent has tools for writing files, executing code, reading output, and searching the web for documentation.
Implemented sandboxed execution with Docker. Each code run happens in an isolated container with resource limits. No access to host filesystem or network unless explicitly granted.
Added artifact storage for generated outputs: code files, CSVs, images, HTML pages. Each artifact is versioned and can be previewed inline.
Created a live preview system for web outputs. HTML/CSS/JS artifacts render immediately in an iframe alongside the chat.
Outcome
The agent successfully completes multi-step tasks that would stump generation-only systems: "Build a chart of this CSV data" requires reading the file, writing visualization code, running it, and fixing any errors. The sandbox execution makes this safe and reliable.