What it does

lisa.py wires LLDB into the Model Context Protocol (MCP), the same standard Claude Desktop uses for tool calling. One half runs as an MCP server; the other loads as an LLDB plugin speaking JSON-RPC. The result: an AI assistant can create targets, set breakpoints, read memory, evaluate expressions, and run arbitrary LLDB commands without you typing a thing.

The interesting bit

The split architecture is the clever part. The plugin lives inside LLDB’s Python runtime and exposes debugger state via HTTP on localhost:13338. The MCP server translates Claude’s natural-language requests into structured JSON-RPC calls. This keeps the heavy debugging logic in-process while letting the AI client stay blissfully ignorant of LLDB’s command syntax.

Key highlights

• Full debugging vocabulary exposed to LLMs: breakpoints, backtraces, disassembly, memory reads, expression evaluation, process attach/detach
• Works two ways: Claude Desktop integration via fastmcp, or direct LLDB use with mcp start
• Auto-load via ~/.lldbinit for the direct path
• Includes run_lldb_command as an escape hatch for anything not in the structured API
• Credits a who’s-who of reverse-engineering tooling (voltron, gef, chisel, ida-pro-mcp) as prior art

Caveats

• README still says “clone lldb-mcp” while the repo is named lisa.py; the install paths don’t quite match
• TODO lists “update instruction manuals” and “add more testcases” as incomplete
• Port 13338 is hardcoded; collision with another service means manual conflict resolution

Verdict

Worth a look if you debug on macOS/ARM64 regularly and already use Claude Desktop. Skip it if you want a polished, fully-documented product today — this is functional glue code with rough edges, not a shrink-wrapped tool.