CodeBurn: free, local-first cost tracker for 31 AI coding tools

June 29, 2026·AI Newsroom·4 min read·10 sources
codeburnai-codingcost-trackingtoken-spendlocal-firstclaude-code+22
CodeBurn TUI dashboard showing a week of AI coding spend broken down by tool, by model, by project, by activity, by day, by core tools, by shell commands, and by MCP server.
Source: github.com/getagentseal/codeburn/blob/main/assets/dashboard.jpg · Captured 2026-06-29 via Playwright Chromium + @sparticuz/chromium (re-rendered with editorial caption and license) · License: MIT (project); screenshot used editorially for the CodeBurn article on news.lesbass.com

CodeBurn is a free, MIT-licensed, local-first TypeScript CLI that reads the session files your AI coding tools already write to disk and breaks down every token and dollar by task, model, tool, and project — across 31 integrations, with no proxy, no API key, and nothing leaving your machine. As of 2026-06-29 the repo sits at 8,313 stars, 650 forks, 769 commits on main, 16 open issues, 15 open PRs (GitHub repository, 2026-06-29). The current line is v0.9.14 (2026-06-22) (release notes); the macOS menubar app ships separately as mac-v0.9.14 on the same date. This is a builder-focused tool piece, not a benchmark.

What it does

The local-first architecture. Per the README, “Everything runs locally. No wrapper, no proxy, no API keys, nothing leaves your machine” (README, 2026-06-29). CodeBurn reads session files the tools already write — Claude Code’s ~/.claude/projects/<sanitized-path>/<session-id>.jsonl, Codex’s ~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl, Cursor’s globalStorage/state.vscdb, OpenCode’s ~/.local/share/opencode/opencode*.db, Gemini CLI’s session files — and joins them by timestamp. Pricing comes from LiteLLM cached locally for 24 hours, with hardcoded fallbacks for Claude and GPT-5 models; v0.9.12 added automatic gap-fill from models.dev and OpenRouter so Fable 5 and Mythos 5 price correctly.

The TUI dashboard. codeburn opens an interactive TUI for the last 7 days. The header shows cost, calls, sessions, in / out / cache-write / cache-read tokens, and cache hit rate. Below: Daily Activity, By Project, By Activity (13 deterministic categories), By Model, Core Tools (Bash, Read, Edit, Grep, Write), Shell Commands, and MCP Servers. The dashboard auto-refreshes every 30 seconds.

v0.9.14 additions (2026-06-22). A new codeburn web browser dashboard (React + Recharts) on localhost:4747. PIN-paired multi-device combining via codeburn share --pair. New providers: Grok Build, ZCode / z.ai GLM-5.2, Hermes Agent, Kiro CLI, Zerostack. A codeburn overview plain-text monthly summary, Codex credit usage, MCP server usage in JSON / CSV exports, and OpenCode 1.1+ file-based sessions. v0.9.12 added codeburn mcp server (get_usage, get_savings over stdio) and providers Devin, Antigravity IDE, JetBrains via Copilot, coder/mux, Vercel AI Gateway.

The macOS menubar app. mac-v0.9.14 ships native Swift / SwiftUI, refreshes every 30 seconds, supports Today / Week / Month / 6 Months, and v0.9.14 added a custom daily-budget alert.

codeburn optimize — the waste-finder. Scans the last 30 days for known cost leaks: re-read files, low Read:Edit ratio, uncapped BASH_MAX_OUTPUT_LENGTH, unused MCP servers, ghost agents/skills/slash commands, bloated CLAUDE.md files, cache-creation overhead, possibly low-worth expensive sessions. Each finding shows estimated savings plus a copy-pasteable fix. Findings are ranked by urgency and rolled up to an A–F setup-health grade.

codeburn yield — the productivity breakdown. Correlates AI sessions with git commits and classifies spend as Productive (commits landed in main), Reverted (commits were later reverted), or Abandoned (no commits near the session). Requires a git repository.

Currency and plans. 162 ISO 4217 currencies (FX from the European Central Bank via Frankfurter, refreshed every 24 hours, no API key). Subscription-plan tracking for Claude Pro / Max, Cursor Pro, and per-provider custom plans via codeburn plan set.

Why it matters

Five reasons this is a day-one story on 2026-06-29.

  1. AI coding spend is the silent line item. Most builders running Claude Code, Codex, and Cursor simultaneously have a spend problem they cannot see. CodeBurn is the only free, local-first, 31-tool option the Writer can verify end-to-end.
  2. Local-first is the differentiator. Tools that cut tokens already exist. CodeBurn is about seeing the bill before you cut it.
  3. 31 integrations is the breadth argument. A team using Claude Code, Codex, OpenCode, and Gemini CLI simultaneously can see one combined total.
  4. The waste-finding loop is the workflow story. codeburn optimize does not just report spend; it suggests copy-pasteable fixes.
  5. The yield command is the executive summary. Productive vs Reverted vs Abandoned spend, correlated with git commits, is the question every AI-leadership meeting asks.

Practical implications

Risks and caveats

  1. Single-maintainer project. @iamtoruk is the primary maintainer; v0.9.14 credits 8 external contributors, but the project is still 1.0 of one person.
  2. Token counts are accurate only for tools that write real counts. GitHub Copilot legacy CLI, Kiro, and Mistral Vibe are estimated — Copilot editor transcripts carry no explicit counts; Kiro sessions are labelled kiro-auto and costed at Sonnet rates.
  3. Pricing accuracy depends on the upstream model name matching the LiteLLM catalog. Behind a proxy that rewrites model names, CodeBurn will report $0.00 until the user configures codeburn model-alias.
  4. optimize findings are heuristic, not causal. Low Read:Edit ratio implies retry loops is a pattern, not a measurement of waste.
  5. Yield requires a git repository. Sessions outside a git working directory are misclassified as Abandoned.
  6. No head-to-head benchmark against an alternative spend tracker is on the record as of 2026-06-29.

What to watch

Sources