Decoder parses the Python AST to build a call graph stored in a local SQLite file, then lets you trace full call chains and see execution context (conditionals, loops, try/except). I built this first as a VS Code extension, but saw the value in giving LLMs that same visibility and added an MCP server. Instead of iterative grep and file reads, an LLM can traverse the call graph directly — which cuts down on token usage and back-and-forth significantly.