apwn/zubo/code-interpreter

code_interpreter

Execute Python, JavaScript, or TypeScript code in a persistent sandboxed workspace and return the output. Use this to run calculations, data processing, text manipulation, chart generation, or any code the user requests.

Code runs inside ~/.zubo/workspace/code-interpreter/ and files are not deleted after execution, so scripts can reference outputs from previous runs.

Input Schema

{
  "type": "object",
  "properties": {
    "code": {
      "type": "string",
      "description": "The code to execute"
    },
    "language": {
      "type": "string",
      "enum": ["python", "javascript", "typescript"],
      "description": "Programming language to use"
    },
    "packages": {
      "type": "array",
      "items": { "type": "string" },
      "description": "Optional list of packages to install before execution. For Python these are installed via pip; for JS/TS via bun add."
    }
  },
  "required": ["code", "language"]
}

Output Schema

The tool returns a JSON string with the following fields:

Usage Hints

• Use "javascript" or "typescript" for JS/TS code (runs via Bun).
• Use "python" for Python code (requires python3 installed).
• Code runs in a persistent workspace (~/.zubo/workspace/code-interpreter/), so files written by one execution are available in subsequent executions.
• The working directory is set to the workspace, so relative file paths work naturally (e.g., open("data.csv") or savefig("chart.png")).
• To install packages, pass them in the packages array:
• Python example: "packages": ["pandas", "matplotlib"]
• JS/TS example: "packages": ["lodash", "chart.js"]
• When your code saves charts or images, the files array in the response will contain the absolute paths to the newly created files.
• stdout and stderr are captured and returned; large outputs are truncated.
• Dangerous system commands (rm -rf /, shutdown, etc.) are blocked.
• Package install has a 60 s timeout; code execution has a 30 s timeout.