April 16, 2026
Tokens on fire, devs on edge
Show HN: CodeBurn – Analyze Claude Code token usage by task
New tracker exposes AI coding costs — devs scream vs shrug
TLDR: CodeBurn is a local dashboard that shows where your Claude Code tokens — and money — go by task and project. Comments exploded over a claimed $1.4k/week spend versus “$200/month is fine,” plus gripes about Cursor support and shout‑outs to open-source alternatives, fueling a broader trust-and-transparency debate.
Meet CodeBurn, the “where did my AI money go?” app that reads your local Claude Code logs and shows exactly what tasks burned your tokens — with a flame icon taunting you from the Mac menu bar. The demo dropped, and the comments immediately lit up like a budgeting bonfire.
The flashpoint? One user quoted the maker’s claim of spending about $1,400 a week on Claude Code, and the thread split into two camps: the “my wallet is melting” crowd vs the “$200/month and never hit limits” skeptics. It’s part cost horror story, part humblebrag, part “do you even refactor, bro?” energy — and everyone’s doing mental math.
Meanwhile, tinkerers are vibing with the terminal dashboard. One commenter loved the Ink (React-for-terminals) look — “just like Claude Code” — then casually dropped, “I don’t trust existing harnesses, so I’m building my own.” DIY paranoia is the new productivity. Others cheered the no-cloud, no-API-key approach as refreshingly privacy-friendly.
But drama’s not done: compatibility complaints rolled in (“doesn’t work with Cursor Agent”), and rival tools surfaced. Fans linked Claudoscope and the fully open-source ClaudeRank, turning the comments into a feature face-off. Verdict? CodeBurn’s slick, local, and spicy — but unless it supports more editors, that menu-bar flame might keep roasting the same crowd.
Key Points
- •CodeBurn is a CLI/TUI tool that analyzes Claude Code token usage by task, tool, model, MCP server, and project.
- •It reads local Claude Code JSONL transcripts directly, requiring no wrappers, proxies, or API keys, and deduplicates by API message ID.
- •Costs are calculated using LiteLLM pricing (cached 24h) with support for input/output/cache read/write and fast mode multipliers, with hardcoded fallbacks.
- •Features include an interactive dashboard, macOS menu bar widget via SwiftBar, and CSV/JSON export with commands for Today/7 Days/Month.
- •The classifier deterministically assigns 13 task categories based on tool usage and keywords, and the tool provides per-model breakdowns (e.g., Opus, Sonnet, Haiku, GPT-4o, Gemini).