Developer Creates Local Token Analytics Tool After $1,400 Weekly Claude Code Spending
AgentSeal has released CodeBurn, an open-source terminal UI dashboard for analyzing AI coding assistant token usage, after discovering limited visibility into nearly $1,400 in weekly Claude Code expenses. Announced via Show HN on April 13, 2026, the project gained 2,276 GitHub stars in three days and received 70 points with 15 comments.
The creator explained: "Built this after realizing I was spending ~$1400/week on Claude Code with almost no visibility into what was actually consuming tokens. Tools like ccusage give a cost breakdown per model and per day, but I wanted to understand usage at the task level."
Analysis Reveals 56% of Spending Goes to Conversation with No Tool Usage
CodeBurn reads JSONL session transcripts that Claude Code stores locally in ~/.claude/projects/ and classifies each turn into 13 categories based on tool usage patterns. The classification is fully deterministic with no LLM calls involved. Analysis of the creator's usage revealed that approximately 56% of spending went to conversation turns with no tool usage, while actual coding (edits/writes) represented only 21% of token consumption.
The tool operates entirely locally with no network requests, telemetry, wrappers, proxies, or API keys required. It supports Claude Code, Codex CLI, Cursor, and Gemini CLI, with multi-currency display capabilities.
Interactive Terminal Interface Provides Task-Level Cost Breakdowns and Success Rates
CodeBurn features an interactive terminal UI built with Ink (React for terminals), offering gradient bar charts, responsive panels, and keyboard navigation. The dashboard provides cost breakdowns by project, model, and activity type, along with daily burn charts, tool and MCP server usage tracking, and cache hit rates.
A unique feature is the one-shot success rate per activity type, which shows where AI succeeds on the first try versus entering retry loops. This visibility helps developers optimize workflows and identify wasteful patterns—a capability missing from existing tools that only show cost per day or model.
Tool Addresses Critical Gap in AI Coding Observability
CodeBurn is published on npm as 'codeburn' and can be installed globally with npm install -g codeburn, then run by typing 'codeburn' in any terminal. The project includes SwiftBar menu bar integration for macOS users.
The tool fills a significant observability gap in AI-assisted development. While most users have visibility into daily costs per model, they lack breakdowns by task type or success rate, making it difficult to optimize workflows or identify inefficient patterns in their AI coding assistant usage.
Key Takeaways
- CodeBurn analyzes AI coding assistant token usage at the task level, revealing that 56% of spending goes to conversation with no tool usage
- The open-source tool gained 2,276 GitHub stars in three days after addressing a $1,400/week visibility problem
- Classification is fully deterministic with no LLM calls, network requests, or telemetry—everything runs locally
- One-shot success rates per activity type help identify where AI succeeds immediately versus entering costly retry loops
- Supports Claude Code, Codex CLI, Cursor, and Gemini CLI with multi-currency display and interactive terminal UI