There's a GitHub repo that racked up over nine thousand stars in under four months, primary language TypeScript. It doesn't train models, build editors, or sell cloud services. It does exactly one thing: show you where your AI coding spend actually goes. This is getagentseal/codeburn. As of August 2026, the repo has 9,092 stars, TypeScript-dominant, MIT license, created April 13, 2026, desktop version at v0.9.19. The npm package is codeburn, requiring Node 22.13+. The README opens with one sentence: "See where your AI spend goes." It's a free, open-source, local-first tool that tracks the token usage and cost of your AI coding tools, across 36 tools and agents, broken down by model, project, task, and tool. The organization behind it is AgentSeal (agentseal.org), and it's a Claude / Codex Open Source Recipient.
What Problem It Solves: The Bill Shows a Total, Never Where It Went
The problem is concrete. You pay for Claude, Codex, Cursor, and a stack of other AI tools. The monthly bill gives you a single total. It never tells you that half of it went to conversation instead of code, that an expensive model burned your budget on work a cheaper one would have one-shot, or that the same file got re-read across sessions over and over. Each tool has its own usage page, in a different format, on a different cadence, and you bounce between them unable to assemble a single coherent picture. CodeBurn fixes exactly this. It reads the session files those tools already write to disk and breaks down every token and every dollar by task, model, tool, and project, so you can see where the money actually went. In the README's words: "You pay for Claude, Codex, Cursor, and a stack of other AI tools. The bill tells you the total. It never tells you that half of it went to conversation instead of code."
Core Mechanism: It Reads Local Files, It Doesn't Sit in the Path
Where CodeBurn diverges from alternatives is the very first step -- how it gets the data. It's not a proxy. It doesn't intercept your API traffic. It doesn't ask for your API keys. Nothing leaves your machine. It reads the session files your AI tools already write to disk: Claude Code's jsonl under ~/.claude/projects/, Codex's ~/.codex/sessions/, Cursor's SQLite state.vscdb, OpenCode's opencode.db, Gemini CLI's session json under ~/.gemini/tmp/, and so on. Each provider is one parser file; adding a new tool means adding one file (the README points at src/providers/codex.ts as the template). Pricing comes from the open-source LiteLLM project, refreshed every 24 hours and cached at ~/.cache/codeburn/, with hardcoded fallbacks for every Claude and GPT-5 model so a fuzzy model-name match never silently prices out to $0.
Four surfaces, one source of truth: a desktop app (Desktop), a browser dashboard (Web), a terminal TUI (Terminal), and a macOS menu bar app (Menubar). The README's phrasing: "Four surfaces, one source of truth: everything reads the session files already on your disk." Desktop v0.9.19 ships a macOS Apple Silicon .dmg, an Intel .dmg, a Windows build on the Microsoft Store, and Linux .deb / .rpm / AppImage packages.
36 Integrations
The project claims 36 integrations. The README's provider-logo section explicitly lists 34, and the data-location table adds a few more (Codebuff, formerly Manicode), so the coverage is broad. The ones named in the README include: Claude Code & Claude Desktop, Cline, CodeWhale, Codex (OpenAI), Cursor, cursor-agent, Devin, Forge, Gemini CLI, Mistral Vibe, GitHub Copilot, IBM Bob, Kiro, OpenCode, OpenClaw, Pi, OMP (Oh My Pi), Droid, Roo Code, KiloCode, Qwen, Kimi Code CLI, LingTai TUI, Goose, Antigravity, Crush, Warp, Mux (coder), Vercel AI Gateway, Zerostack, Grok Build, ZCode, Zed, and Hermes Agent. Provider data locations and formats vary widely: Cursor lives in SQLite, Codex in date-bucketed jsonl, GitHub Copilot prefers an OTel SQLite store, and Zed stores zstd-compressed blobs (needing Node 22.15+'s built-in zstd). CodeBurn normalizes all of it into one table, split by provider / model / project / task. The --provider flag restricts any command to a single tool.
It Doesn't Just Show Waste, It Fixes It
codeburn optimize scans the last 30 days of sessions plus your ~/.claude/ setup for waste patterns: files Claude re-reads across sessions, a low Read:Edit ratio (editing without reading leads to retries and wasted tokens), uncapped bash output, MCP servers still paying tool-schema overhead every session but never invoked, ghost agents / skills / slash commands defined but never triggered, bloated CLAUDE.md files (with @-import expansion counted), cache-creation overhead and junk-directory reads, context-heavy sessions where input swamps output, and low-worth expensive sessions with no edit turns and no git delivery. Each finding shows the estimated token and dollar savings plus a ready-to-paste fix -- a CLAUDE.md line, an environment variable, or a mv to archive unused items. Findings are ranked by urgency and rolled into an A-to-F setup-health grade.
The sharper move is codeburn optimize --apply, which applies config-class fixes for you: settings values, environment variables, archiving unused agents and skills. Every change is backed up and journaled before it lands, and codeburn act undo rolls it back. And it closes an honesty loop: once an applied fix is at least three days old, codeburn act report compares its estimated savings against what your sessions actually did. Estimates get checked against reality, not just claimed.
Putting a Lock on the Budget
codeburn guard install drops opt-in hooks into Claude Code's .claude/settings.json that watch session cost while you work: a soft cap (default $5) that warns once mid-session, a hard cap (default $15) that stops the session (lift it for that session with codeburn guard allow), and a checkpoint (default $3) that, if a session ends past it with no edits and no commits, nudges you to start fresh with a named deliverable. The hooks fail open -- a broken guard never blocks your session. Caps are edited in ~/.config/codeburn/guard.json.
Model Comparison, and "Did the Spend Actually Ship"
codeburn compare runs a model head-to-head on your own real data: one-shot rate (edits that succeed without retries), retry rate, self-correction, cost per call, cost per edit, cache hit rate. It answers "which model is actually better for my work," not a benchmark leaderboard.
codeburn yield correlates AI sessions with git commits by timestamp: Productive (this session's commits landed in main), Reverted (commits were later reverted), Abandoned (no commits near the session, or commits never merged), Ambiguous (the session ran parallel to another and its window's commits were attributed to the tighter one). It answers "I spent this much -- did it turn into code?"
Installation: npx One-Liner, or Platform Downloads
The fastest path, no install:
npx codeburnThis opens the interactive TUI dashboard, last 7 days by default. Arrow keys switch periods, q quits. In 30 seconds you know where your AI budget went.
Install it as a permanent command:
npm install -g codeburnAlso available via bunx codeburn, pnpm dlx codeburn, and brew install codeburn on macOS. Requires Node.js 22.13+, and at least one supported tool with session data on disk. Cursor and OpenCode auto-install better-sqlite3.
The desktop app downloads straight from GitHub Releases: macOS Apple Silicon (CodeBurn-0.9.19-arm64.dmg), macOS Intel (CodeBurn-0.9.19.dmg), Windows via the Microsoft Store, Linux as .deb / .rpm / AppImage. The macOS menu bar app is one command -- codeburn menubar -- which downloads the latest .app, installs it into ~/Applications, and launches. Linux gets the same ambient view through a GNOME Shell extension (GNOME 45+). On Windows, codeburn web is the always-on view for now.
Local browser dashboard: codeburn web opens a charted dashboard at http://localhost:4747. The server binds to localhost only; nothing is uploaded.
Combine usage across devices: codeburn share --pair on another device shares its usage (printing a PIN), then codeburn devices add on your main device pairs it. PIN-authorized, stays on your local network.
For your agent (MCP): claude mcp add codeburn -- npx -y codeburn mcp starts a local stdio MCP server exposing two tools -- get_usage (spend and usage broken down by tool, model, project, and task; fast) and get_savings (cost reductions: waste findings, retry tax, routing waste; slower, deeper). Project names are pseudonymized by default; the agent only sees real names if it asks with include_project_names: true.
Versus the Alternatives: Vendor Usage Dashboards and Cloud Bills
The most direct "alternative" is each vendor's own usage page. Claude Console, OpenAI's usage, and Cursor's admin console all give you a total and a per-model breakdown, but they share the same shortcomings. First, each only covers itself -- use three and you're opening three tabs and adding them up by hand. Second, the granularity stops at "model / day," never reaching "task / project," and never tells you whether a session was conversation or code. Third, a cloud bill covers what it bills you for; when you're paying via subscription quota (Claude Pro / Max, Cursor Pro), it won't break your real token consumption out for you.
CodeBurn's differences are three. One, 36 providers in a single table with a unified schema. Two, a task-classification system (13 categories, determined purely from tool-usage patterns and user-message keywords -- no LLM calls, fully deterministic) that separates "conversation / exploration / coding / debugging," surfacing conclusions no bill will show you, like "half your spend went to conversation." Three, yield ties spend to git commits, answering "did the money turn into code," which no cloud bill does. The trade-off: it can only track tools it integrates with, and it depends on those tools writing sessions to disk -- anything not integrated is invisible, and when a vendor changes its session format, CodeBurn has to follow (the README itself says "keeping 36 integrations honest is constant work; the tools underneath change every week").
Compared with cloud-based cost-monitoring services, CodeBurn's structural advantage is "local-first": your code, prompts, and sessions never leave the machine, and the only thing pulled from the network is the LiteLLM price file, cached locally. For teams that care about privacy or can't ship sessions to a third party, that's a structural win.
Echo From This Project
A genuine connection: this very site (aiwebcool.com) is built with coding agents like Claude Code -- writing articles, changing code, deploying, all done in conversation with an agent. We've lived the exact failure mode CodeBurn describes: a session burns half a day of tokens and produces barely any commits, and the bill shows one total with no way to tell which stretches of the conversation were real work versus the agent spinning its wheels. "See where your AI spend goes" isn't a nice-to-have for a one-person company running an AI content site on agents -- it's directly whether the monthly cost makes sense. What a tool like CodeBurn does is turn "AI coding bills are opaque" from something you endure into something you can see.
Who It's For, and What to Watch For
Suitable for: developers using multiple AI coding tools who want one consolidated ledger; anyone paying via subscription quota (Pro / Max) who wants to know whether their real token consumption is within the plan; tech leads running teams on coding agents who want to put a hard cap on session cost; teams that care about privacy and can't upload sessions to a third-party monitoring service.
Five caveats. First, it needs Node 22.13+ (the README is explicit), and the Zed provider specifically needs Node 22.15+ for built-in zstd; an older Node and it won't run. Second, it only tracks tools it integrates with -- your AI tool has to be among the 36 integrations and has to write sessions to a path CodeBurn can find, otherwise it's zero. codeburn doctor diagnoses each provider: which paths it probed, how many sessions it found, parse health. Third, data precision varies a lot across providers: Claude Code, Codex, and Gemini CLI carry real token counts; Cursor's output is estimated from reply text, cache is server-side and uncountable, and long conversations get undercounted -- the README itself marks these "estimated"; Kiro doesn't expose the model, prices at Sonnet rates, and estimates tokens from content length. Reading the numbers, distinguish "measured" from "estimated." Fourth, optimize --apply modifies your config files (backed up and undoable, but still), so run --dry-run first on anything production-facing. Fifth, the guard hooks are explicitly written for Claude Code -- it can't enforce a hard cap on other agents. Sync (team telemetry) is still in preview; the protocol may change between releases.
The People Behind It and the Model
Behind it is AgentSeal (agentseal.org), a Claude / Codex Open Source Recipient (the badge links to claude.com/open-source-max). Community lives on Discord, the X account is @_codeburn, and GitHub Sponsorship goes through iamtoruk. The business model is straightforward: the tool itself is MIT, free, and runs locally, funded by GitHub sponsorship -- the README's phrasing is "if it has already saved you more than a sponsorship costs, consider sending a little of that back." It's explicit that maintaining 36 integrations is constant work (vendors change every week -- Cursor reshapes its database, Claude moves a config path, new models ship at new prices), and sponsorship pays to "keep the numbers you see always the real ones." Team or company sponsors get their logo in front of every developer who opens the repo. Note that the project explicitly states it's not affiliated with CodeBurn Bt. or codeburn.hu.
CodeBurn isn't complicated -- it has no model code and provides no cloud service. What it does is read the session files already on your machine, normalize them, and split them by task, model, and project, turning the "AI coding bill" from a monthly total into a ledger that shows where the money went. Behind those nine thousand stars are people who got tired of being shown a total and decided that seeing where the spend goes is, by itself, worth it.
References
- CodeBurn GitHub repo (9,092 stars, TypeScript, MIT license): https://github.com/getagentseal/codeburn
- README (positioning, 36 integrations, command list): https://github.com/getagentseal/codeburn/blob/main/README.md
- npm package codeburn: https://www.npmjs.com/package/codeburn
- AgentSeal site: https://agentseal.org
- Open Source Recipient page (Claude / Codex): https://claude.com/open-source-max
- Pricing source LiteLLM: https://github.com/BerriAI/litellm
- GitHub Sponsor (iamtoruk): https://github.com/sponsors/iamtoruk