CodeBurn is an open-source cost and token observability tool for AI coding agents built by AgentSeal. It ships as a Node CLI that opens an interactive TUI dashboard showing exactly where your AI coding tokens go: broken down by task type, tool, model, MCP server, and project. Instead of a wrapper, proxy, or SDK, CodeBurn reads session data directly off disk from each supported agent — Claude Code (`~/.claude/projects/`), Codex (`~/.codex/sessions/`), Cursor, OpenCode, Pi, and GitHub Copilot — so there are no API keys to configure and nothing to install in your IDE.
The dashboard is built around gradient charts, responsive panels, and keyboard navigation, with a provider plugin system so new agents can be dropped in without touching the core. Beyond raw token and dollar totals, CodeBurn tracks the one-shot success rate per activity type, surfacing the gap between tasks where the AI gets it right first try and tasks where it burns tokens cycling through edit/test/fix loops. Pricing is pulled from LiteLLM and auto-cached, covering every supported model, with configurable currency and a macOS menu bar widget via SwiftBar for always-on visibility.
CodeBurn is MIT licensed and installs with `npm install -g codeburn` or runs directly with `npx codeburn` on Node 20+. Data never leaves your machine; the tool simply parses local session files and produces CSV/JSON exports when you need to pipe numbers into a spreadsheet or team dashboard. For solo developers it answers "am I getting value from my $20/month plan?" and for engineering managers it quantifies which agents, projects, and task types are quietly becoming the most expensive — a growing need as AI coding spend moves from hobby experiment to real line-item on the cloud bill.
