Does a Language Server Save Tokens for Coding Agents? A Measurement Methodology and Preliminary Study

A preliminary study evaluating Python and TypeScript repositories shows that semantic retrieval via the Language Server Protocol (LSP) often fails to improve token-to-success efficiency compared to lexical retrieval. Using Claude Opus, Sonnet, and Haiku, the researchers find that the token overhead of LSP-driven symbol localization frequently outweighs its precision benefits for coding agents.