BeagleLathe started as a private hack after watching Claude Code spend twelve tool calls renaming handleAuth across three files. Glob, Grep, three Reads, an Edit, a verify-Read, a failed Edit, a retry, another verify, then the third file. I would watch agents burn through tons of tokens and time reading/editing/searching for files. It seems sort of weird that we have these advanced AI systems but the way they access the filesystem is not very optimized. Measured head-to-head on nine fixtures, same model on both sides: ~10% cheaper, ~46% fewer tool calls, ~29% faster. Up to ~19% cheaper on multi-file refactors. 45/45 verify-pass — the savings don't come from cutting corners. It runs on your machine. Source files never go to a server. Are your agents doing any weird tool-call loops? What else do you think could be optimized with more intelligent tool calls.

BeagleLathe

Claude Code, fewer tool calls, faster coding.

Claude Code, fewer tool calls, faster coding.