CodeBurn

The HackerNews discussion (34 comments, all +112) reveals a developer tool that fills a genuine gap: visibility into Claude Code token consumption. Users were surprised by the finding that ~56% of turns are conversation (planning, context management, thinking) vs ~21% actual coding — one user noted this aligns with SWE-bench trajectory studies showing ~38% of agent actions are pure exploration (grep, find, file reads). The JSONL transcript parsing is praised as 'clever' and 'reliable.' Cache hit rate tracking (one user reported 98.2% on Opus 4.6) is highlighted as the difference between a $2K month and something far worse. Activity classification is called 'the actually useful feature' vs. simple spend trackers. However, accuracy concerns exist: one user reports only '1 turn of planning' detected in 30 days despite extensive planning workflows, suggesting the classifier may miss or mislabel activities. The tool currently reads the Cursor IDE's state.vscdb but does NOT yet parse cursor-agent CLI transcripts (~/.cursor/projects/*/agent-transcripts/) — the developer acknowledged this and filed an issue to add it. The tool is built with Ink (React for terminals) and requires no LLM to operate. A side debate emerged about cost context: the creator clarified the ~$1,400/week figure is the API-equivalent token cost, not out-of-pocket spend — they are on the $200/month Max plan. Users on $200 subscriptions report rarely hitting rate limits even with 5+ concurrent agent tabs on 300K LoC codebases, while API users at $1,400/week get unthrottled parallel execution. NOTE: The Reddit comments in the dataset are entirely unrelated (Tesla, Kia, Ford cars, AI-generated article complaints) and provide zero signal about CodeBurn.

No aggregate ratings were found for this product during the last harvest.