CodeBurn

Core Positioning#

CodeBurn is a fully local cost observability dashboard for AI coding tools. When developers use multiple AI coding tools simultaneously, there's a lack of transparent correlation between token consumption and actual output. CodeBurn reads session logs/databases directly from local disk — zero intrusion, zero configuration for full cost visibility. It only performs "read + analyze + display" — no interception, no proxying, no modification of any AI tool behavior.

Session Data Collection#

Covers 18 AI coding tools: Claude Code, Claude Desktop, Codex (OpenAI), Cursor (SQLite database reading), cursor-agent, Gemini CLI, GitHub Copilot (CLI + VS Code workspaceStorage), Kiro, OpenCode, OpenClaw, Pi, OMP, Droid, Roo Code, KiloCode, Qwen, Goose, Antigravity. Pure local disk reading — no wrapper/proxy/API key required.

Cost Calculation Engine#

Based on LiteLLM model pricing data with 24-hour local cache. Itemized calculation of input / output / cache read / cache write / web search token costs. Supports Claude Fast Mode multiplier. Built-in hardcoded fallback pricing to prevent fuzzy matching errors.

Efficiency Metrics#

• One-Shot Rate: Detects Edit/Bash/Edit retry patterns, measures first-edit success rate
• Yield Correlation: Associates AI sessions with git commits by timestamp, classified as Productive / Reverted / Abandoned
• Task Classification: 13 deterministic categories (Coding, Debugging, Feature Dev, Refactoring, Testing, Exploration, Planning, Delegation, Git Ops, Build/Deploy, Brainstorming, Conversation, General) — no LLM calls

Waste Detection (Optimize)#

Scans sessions and ~/.claude/ config to identify cross-session duplicate file reads, low Read:Edit ratios, unused MCP server overhead, bloated CLAUDE.md, ghost agents/skills/slash commands. Outputs A-F health rating + paste-ready fix suggestions.

Model Comparison (Compare)#

Interactive panel comparing one-shot rate, retry rate, self-correction, cost per call, cache hit rate and other metrics.

Budget & Internationalization#

Subscription plan tracking: Claude Pro/Max, Cursor Pro presets + custom monthly budget with dashboard progress bars. Supports 162 ISO 4217 currencies with exchange rates from Frankfurter (European Central Bank data).

Desktop Integration & Data Export#

macOS Menu Bar app (codeburn menubar), Linux desktop integration (gnome/ directory). --format json for all report commands for pipeline processing, codeburn export for CSV export.

Architecture Overview#

TypeScript ESM modules, tsup bundling, Commander.js CLI framework, Ink (React for CLI rendering) + React 19, Vitest testing. Provider architecture uses single-file design (e.g., src/providers/codex.ts) for easy extensibility. Integrates Semgrep security rules. Source structure: src/ (core), tests/, scripts/, mac/ (menu bar app), gnome/ (Linux desktop).

Installation & Usage#

Prerequisites: Node.js ≥ 22.20 (macOS / Linux / Windows), at least one supported AI coding tool with local session data.

npm install -g codeburn
brew tap getagentseal/codeburn && brew install codeburn
npx codeburn

Core commands: codeburn (interactive dashboard), codeburn today/month, codeburn report -p 30days, codeburn status, codeburn optimize, codeburn compare, codeburn yield, codeburn export, codeburn plan set claude-max, codeburn currency GBP. Config file: ~/.config/codeburn/config.json.

Unconfirmed Information#

• Exact NPM package URL (README provides install command but no direct npmjs.com link)
• AgentSeal organization background is limited (only email hello@agentseal.org), no independent website
• Menu Bar app code signing and notarization status unspecified
• Current version 0.9.7, not yet 1.0 — CLI interface may have breaking changes
• Long-term compatibility of Cursor SQLite reading format remains to be observed