Let Your LLM Actually Run the Debugger

What it does This project is a VS Code extension paired with an MCP server that hands your IDE’s debugger over to an LLM. Once connected, the model can set breakpoints, evaluate expressions in the debug console, and step through execution to diagnose bugs interactively. It is language-agnostic, relying on your existing launch.json and whatever debugger VS Code already supports for that stack.

The interesting bit Instead of asking an LLM to statically analyze source code and hope it spots the error, you can ask it to observe the program while it runs. The extension exposes live debugging actions as MCP tools, so models like Claude can treat your runtime state as just another context source.

Key highlights

• Works with any language VS Code can debug, provided you have a working debug configuration
• Supports both stdio and SSE transport for MCP clients like Claude Desktop, Continue, and Cursor
• Lets LLMs evaluate arbitrary expressions in the debugger console during a session
• Gracefully hands off the debug server when you switch between multiple VS Code windows
• Already supports conditional breakpoints, with planned tools for auto-fix suggestions

Caveats

• Requires a valid .vscode/launch.json; the LLM cannot bootstrap a debug environment from scratch
• The README notes that file discovery currently lists files rather than using ripgrep, which can be inefficient
• Only explicitly demonstrated with Claude Desktop and Continue, though the MCP standard should allow broader compatibility

Verdict Worth a look if you already debug in VS Code and want an LLM to do the tedious stepping-and-watching work. Skip it if you don’t have a working launch.json or prefer your debugger to stay under human control.