Developer Audits 87 Claude Code Sessions to Trace AI Coding Costs

On November 2, 2026, a developer published an article on DEV Community titled "Recording 87 Claude Code sessions to work out where the money goes." The stated motivation was to find out why AI coding agents get expensive on real codebases, so rather than keep guessing the author started recording. Every session in the article is a real Claude Code run on Claude Sonnet 5, saved as a full transcript together with the cost reported by Claude Code itself. In total there are 87 sessions, all run against two open-source repositories. Along the way the developer wrote knock, an open-source MCP server that hands the agent a map of the codebase. According to the article, the first two versions of knock saved nothing, and the piece walks through what changed once they finally did start to help. That is where the account currently stops: the cost figures from the 87 recorded sessions, the release of knock as an open-source project, and the author's explanation of why the early versions failed to cut costs and what made the difference later on.