mcp-app-python

The Python server code (both server.py and server-html.py) appears robust, handling file reads and HTTP requests without obvious direct vulnerabilities like `eval` or command injection. The `load_html` function in `server.py` involves string manipulation to inline JS/CSS and move script tags for self-contained widget embedding; while this carries an inherent, albeit small, risk if not perfectly implemented, it's for known build artifacts. Client-side widgets are expected to read `window.openai.toolOutput`; developers should ensure proper sanitization/escaping if rendering dynamic content directly as HTML within the widget. The example React widget uses `JSON.stringify` and direct text injection for the 'pizzaTopping', which are generally safe. The client environments (e.g., ChatGPT, MCP Inspector) are expected to sandbox these widgets.