CodeBurn tackles a common challenge for developers using AI-powered coding tools: understanding not just how many tokens you spend, but how efficiently your AI coding assistant works. Beyond simple cost monitoring, CodeBurn surfaces a one-shot success rate metric that reveals how often the AI gets it right on the first try, highlighting wasted tokens spent on retries.
What CodeBurn does and how it works
CodeBurn is a local-first terminal user interface (TUI) dashboard written in TypeScript, running on Node.js 20 or higher. It supports 18 AI coding tools, including well-known models and platforms like Claude Code, Codex, Cursor, and Gemini CLI. Unlike many cloud-based analytics tools, CodeBurn reads session data directly from disk, eliminating the need for proxies, API keys, or external servers.
Under the hood, CodeBurn parses session files produced by the supported AI coding tools, extracting token usage, request metadata, and performance data. It uses LiteLLM to price every API call, caching pricing locally for 24 hours to avoid redundant lookups and speed up the dashboard.
The dashboard organizes data by 13 deterministic task categories derived from observed usage patterns, such as code generation, editing, testing, and debugging tasks. This categorization allows developers to see cost and token usage broken down by task type, model, project, and provider.
One standout feature of CodeBurn is its detection of retry patterns common in AI coding workflows — specifically, the edit → test → fix cycle. It calculates a one-shot success rate metric that measures what percentage of edit turns succeed without requiring retries, giving a tangible metric for AI efficiency.
The tool refreshes its dashboard automatically every 30 seconds by default, reflecting new session data as developers work. It supports global installation via npm or Homebrew and can also be run directly without installation through npx, bunx, or dx.
Technical strengths and tradeoffs
CodeBurn’s architecture prioritizes local-first data processing, which means all token usage and cost analytics happen on the developer’s machine. This approach avoids privacy concerns from sending usage data to third-party servers and reduces setup complexity — no API keys or cloud dependencies are needed.
The codebase is written in TypeScript targeting Node.js 20+, ensuring modern language features and runtime optimizations. The parsing logic handles session files from multiple AI coding tools, which is non-trivial given the variety of formats and update frequencies.
LiteLLM integration for pricing calls is another technical highlight. By caching pricing information locally for 24 hours, CodeBurn minimizes network overhead and provides responsive feedback in the dashboard.
The one-shot success rate metric is technically interesting because it requires detecting retry cycles from session data, a subtle pattern that goes beyond raw token counting. This metric provides actionable insight to developers about AI coding effectiveness, not just cost.
Tradeoffs include the reliance on session data being properly written to disk by the AI tools, so if a tool changes its logging format or storage location, CodeBurn may require updates. Also, the local caching strategy means pricing data could be stale if pricing changes within the 24-hour cache window.
The dashboard is a TUI, which favors terminal-focused developers but might not appeal to those expecting web UIs. However, the refresh every 30 seconds and live breakdowns provide a real-time feel within the terminal.
Quick start
Requirements
- Node.js 20+
- At least one supported AI coding tool with session data on disk
- For Cursor and OpenCode support,
better-sqlite3is installed automatically as an optional dependency
Install
npm install -g codeburn
Or with Homebrew:
brew tap getagentseal/codeburn
brew install codeburn
Or run directly without installing:
npx codeburn
bunx codeburn
dx codeburn
verdict
CodeBurn is well suited for developers who heavily use multiple AI coding tools and want detailed, local insights into token usage and AI coding efficiency. Its one-shot success rate metric is a useful addition that goes beyond cost tracking, helping developers optimize workflows and understand AI performance.
The local-first design limits privacy risks and setup friction but depends on stable session data formats. The TUI approach fits terminal-centric workflows but may not suit everyone.
Overall, CodeBurn is a practical, thoughtfully designed tool that fills a niche in AI coding observability. It’s worth trying if you want to go beyond billing numbers and dig into how effectively your AI coding assistants perform in real development scenarios.
Related Articles
- CC Switch: unified management for AI coding CLIs in a cross-platform Rust desktop app — CC Switch is a Rust-based cross-platform desktop app that centralizes management of AI coding CLIs like Claude Code and
- OpenAI Codex CLI: local-first AI coding assistant with ChatGPT integration — OpenAI Codex CLI brings AI coding assistance local to your terminal, integrating with ChatGPT plans for powerful hybrid
→ GitHub Repo: AgentSeal/codeburn ⭐ 5,212 · TypeScript