Terrarium provides a secure, low-latency Python sandbox for executing untrusted LLM-generated code, targeting developers building data agents. It offers an economical solution for running Python scripts, generating visualizations, and processing data within a compartmentalized environment, with quick execution times and low hosting costs.

How It Works

Terrarium leverages Pyodide, a CPython distribution compiled to WebAssembly, to run Python code within a Node.js process. This approach isolates execution by providing a memory-only filesystem, preventing access to the host's memory or network. Each sandbox instance is completely recycled after every invocation, ensuring no state is carried over between calls.

Quick Start & Requirements

• Install dependencies: npm install
• Run locally: npm run dev
• Execute code: curl -X POST -H "Content-Type: application/json" --url http://localhost:8080 --data-raw '{"code": "1 + 1"}'
• Supported packages: numpy, pandas, matplotlib, sympy, etc. (see Pyodide documentation)
• Deployment: Docker or GCP Cloud Run.

Highlighted Details

• Fast execution: 900ms for PNG generation, 500ms for SVG.
• Economical hosting: <$30/month on GCP Cloud Run (2GB mem, 1vCPU).
• Native file support: Input/output files handled via base64 encoding.
• plt.savefig() supported for Matplotlib; plt.show() is not.

Maintenance & Community

• Developed by Cohere AI.
• Deployment guidance provided for GCP Cloud Run, including health checks.

Licensing & Compatibility

• No license specified in the README.
• Compatibility for commercial use or closed-source linking is not explicitly stated.

Limitations & Caveats

Package installation, network access, and complex computations can be challenging. Pyodide may encounter "Maximum call stack size exceeded" errors with very high DPI settings or complex Pandas operations.