As AI coding assistants become integral to developer workflows, keeping tabs on token consumption and costs across multiple tools — Claude Code, Cursor, Codex, KiloCode, and more — is becoming a real challenge. CodeBurn is a powerful, privacy-first CLI and TUI tool that solves this by reading session data directly from your machine and presenting detailed usage dashboards, without requiring any API keys or proxy servers.

Built with Node.js 20+ and published as an npm package, CodeBurn installs in seconds and works out of the box with a growing list of supported providers.

• 18 Supported AI Coding Tools: From Claude Code and Codex CLI to Cursor, KiloCode, OpenCode, GitHub Copilot, and beyond — CodeBurn reads session data from all of them, giving you a unified view of your entire AI coding ecosystem.
• Local-Only, No API Keys: Everything runs on your machine. CodeBurn parses local session files and prices calls using LiteLLM pricing data. No data ever leaves your machine.
• Rich TUI Dashboard: Interactive terminal dashboard with today’s usage, monthly totals, cost breakdowns, cache hit rates, and a “compare” view. Export to CSV or JSON.
• macOS Menu Bar App: A lightweight menubar widget for real-time token and cost tracking.
• Deep Cursor Support: Handles Cursor’s v3 bubble format, zero token counts, and agentKv fallback paths with iterative community-driven fixes.

The project has an active community with developers reporting real-world bugs. Here are some highlights:

---

Issue #159 — Cursor Cost Tracking Not Working (27 comments)

A user reported that Cursor usage showed zero tokens despite confirmed daily activity. The maintainer @iamtoruk diagnosed two bugs: (1) Cursor’s v3 bubble format stores zero token counts, and the provider filtered on inputTokens > 0, silently dropping everything; (2) the agentKv fallback skipped entries with plain strings instead of JSON arrays. Both were fixed in v0.9.2 and v0.9.3, with the 35-day lookback widened to 180 days.

“The root cause was Cursor’s v3 bubble format storing zero token counts for all entries. The provider was filtering on inputTokens > 0, so everything got dropped. Now it estimates tokens from the text length when the counts are zero.” — @iamtoruk

---

Issue #190 — Is Release v0.9.5 Broken? (16 comments)

After v0.9.5, users reported minimal data — only $0.0031 and 2 calls despite heavy Claude Code and Codex usage. The cause: Codex CLI 0.128+ embeds the full system prompt in the first line, pushing file size past the 16 KB validation read limit. Sessions silently failed validation and were skipped. The fix expanded the file read buffer.

“If you are on Codex CLI 0.128+, this is a known bug where the session validation reads only the first 16 KB of each session file. Newer Codex versions embed the full system prompt in the first line, pushing it past 16 KB, which causes sessions to silently fail validation and get skipped.” — @iamtoruk

---

Discussion #29 — Deno Support as a Security Alternative

A community member raised a security concern: running npx from a young repository that reads local files is a legitimate concern. The proposal to add Deno as an alternative runtime was discussed, highlighting the tension between convenience (npm) and security (Deno’s explicit permissions model).

---

CodeBurn is the first tool that gives developers a single, local, privacy-preserving dashboard for tracking costs and token usage across nearly all major AI coding tools. The active issue tracker and rapid iteration cycle (v0.9.x releases in rapid succession) show a deeply engaged maintainer. For developers and teams using multiple AI coding assistants, CodeBurn is quickly becoming essential.

Install it today: npm install -g codeburn or run with npx codeburn.

@getagentseal / codeburn