Code Burn Tracks Tokens But Lacks Actionable Insights
Code Burn visualizes Cloud Code and Codex usage (e.g., $166 hypothetical cost for Claude), breaking down by project, activity, and tools like bash/PHP—but subscription limits matter more, and Cloud Code's /insights gives optimization tips instead.
Usage Breakdown Drives Visibility Into AI Coding Patterns
Code Burn analyzes Cloud Code and Codex logs via npx codeburn, producing a dashboard with hypothetical API costs based on LiteLLM pricing. For a $100 Anthropic (Claude/Cloud Code) and $20 OpenAI (Codex) subscriber, it showed $166 total Claude cost across projects, prioritizing Cloud Code for primary tasks and Codex for reviews when Cloud Code hallucinates or outages hit. Key metrics include:
- Daily activity: Cost, calls by model (e.g., Opus, GPT-4o), one-shot success rate (edits succeeding without retry cycles like edit-bash-edit).
- By project: Token split across demos/internal work.
- Activity types: Coding, dev, debugging (possibly self-correction attempts).
- Tools: Bash tops (Laravel PHP artisan, MCP calls), followed by read/write/edit files (writes/edits ~300 vs. reads), shell commands (PHP, Git, Biome for styling).
- MCP servers: Laravel Boost primary, Laravel Restify and Context7 as fallbacks.
Toggle providers with 'p' key (Claude vs. OpenAI), filter time (today/month), export CSV/JSON, or run codeburn --provider anthropic --currency USD. Supports custom providers via JSONL parsing; menu bar mode available separately. With 1,000 GitHub stars days post-launch, it echoes CC Usage (log analyzer with recent commits).
Hypothetical Costs Ignore Subscription Constraints
Displayed costs assume pay-per-token (e.g., $166 Claude), but subscriptions enforce 5-hour and weekly limits—untracked here. Colors highlight activity (e.g., green for one-shots), yet they don't guide behavior changes like caching prompts, model switches, or MCP tweaks. No tips emerge from high bash usage or 50% one-shot rates, leaving data inert.
Built-in Insights Outperform Third-Party Dashboards
Cloud Code's /insights command analyzes sessions for patterns and recommendations (e.g., tweak .cursorrules for fewer hallucinations), far more actionable despite high token cost on long histories. Run it post-sessions to optimize prompting/MCPs; demoed in prior coverage of Cursor features. Skip Code Burn's visuals for this native tool to actually reduce retries and boost efficiency.