Show HN: CodeBurn — Analyze Claude Code token usage by task

April 16, 2026
aiprivacystreamingopen-source
Laptop displaying code outdoors, ideal for remote work or freelance programming.
Photo by Meet Patel on Pexels

What it does

Tired of watching your AI coding credits go up in smoke? CodeBurn aims to show exactly where those tokens land — by task type, tool, model, MCP server, and project. The tool presents an interactive TUI dashboard with gradient charts and responsive panels, a macOS menu-bar widget via SwiftBar, and CSV/JSON export so you can slice and dice usage. It supports Claude Code, Codex (OpenAI), Cursor, OpenCode, and Pi through a provider plugin system.

How it works

It has been reported that CodeBurn reads session data directly from disk — no wrapper, no proxy, no API keys required — and aggregates activity into one-shot success rates and tool/model breakdowns so you can see when the AI nails it first try versus burning tokens on edit/test/fix loops. Cursor support reads from a local SQLite DB and, because Cursor’s "Auto" mode hides the model, costs for those runs are estimated using Sonnet pricing (labeled in the dashboard). First run on a large Cursor database may take up to a minute; subsequent runs use cached results.

Install and providers

Install with npm install -g codeburn or run instantly with npx codeburn. Requirements include Node.js 20+ and the presence of local session files (Claude Code: ~/.claude/projects/, Codex: ~/.codex/sessions/, Cursor/OpenCode/Pi paths listed in the repo). The provider plugin system makes adding a new source a single-file task — implement session discovery, JSONL parsing, normalization, and display names and you’re done. By the way, it has been reported that pricing lookups use LiteLLM (auto-cached) and currency conversion supports 162 ISO 4217 codes via an exchange-rate service.

Why it matters

This is the sort of practical tooling teams have been missing: a clear audit trail of where token spend actually happens. Want to blame the model or the workflow? Now you can point fingers with data. It’s on GitHub for inspection and contribution — transparency first, guesswork later.

Sources: github.com/agentseal, Hacker News