Claude Code usage tracking by LangWatch

Product Introduction

1. Definition: LangWatch's Claude Code usage tracking is a specialized OpenTelemetry-based observability and cost analytics platform for AI coding agents. It is a technical monitoring solution that captures, analyzes, and visualizes the complete execution trace of sessions from Claude Code, Codex, OpenCode, Cursor, and other AI-powered coding assistants.
2. Core Value Proposition: It exists to provide developers and engineering teams with granular, actionable insights into their AI coding agent usage, specifically to demystify token consumption, accurately calculate theoretical costs on bundled plans, and enable performance optimization through detailed session replay and trace analysis. Its primary value is in transforming opaque AI agent activity into structured, queryable telemetry data.

Main Features

1. Granular Token & Cost Analytics: LangWatch provides detailed breakdowns of token usage across distinct classes: Input, Output, Cache Read, and Cache Write. It calculates both the theoretical API cost (based on official per-token pricing) and marks sessions as "Bundled" for subscription plans (Claude Pro, Max, Team), showing users exactly what their usage would cost on a pay-per-token model. This works by ingesting native OpenTelemetry data exported directly from the coding agents.
2. Comprehensive Trace & Session Replay: Every interaction is captured as a trace containing nested spans. This includes each model turn (e.g., a call to claude-opus-4-8), every Bash command execution, file edit operation, and MCP (Model Context Protocol) tool call. Each span records duration, input/output data, and associated tokens. The platform provides multiple visualizations (Waterfall, Flame, Topology, Sequence graphs) and a full terminal replay, allowing for deep forensic analysis of agent behavior and bottlenecks.
3. Integrated MCP Server for Agent Self-Analysis: A key technical feature is LangWatch's own MCP server, which exposes the collected trace history to the AI coding agent itself. This allows Claude Code to query its own past sessions, analyze patterns in token expenditure, identify inefficient workflows, and potentially self-optimize its future actions based on historical performance data.
4. Automated PII & Secret Redaction: The platform automatically scans and redacts sensitive information, including API keys and personally identifiable information (PII), from trace data before it is stored. This is a critical security and compliance feature, ensuring that accidental leaks during a coding session do not persist in the observability platform.

Problems Solved

1. Pain Point: Lack of visibility into AI coding agent costs and efficiency, especially for users on fixed-price "Bundled" subscription plans (Claude Pro/Max/Team). Users cannot answer "What is this session costing me?" or "Where are my tokens being spent?"
2. Pain Point: Inability to debug or optimize agent workflows. When a Claude Code session produces an unexpected result or gets stuck, developers have no native tools to replay the session, examine intermediate tool calls, or identify time spent waiting.
3. Target Audience: Software Developers & Engineers who regularly use AI coding assistants for daily tasks; Engineering Managers & Tech Leads responsible for team productivity and software development tool budgets; DevOps & Platform Engineers tasked with implementing observability and cost control for AI tooling across an organization.
4. Use Cases: Cost Attribution and Forecasting for teams to understand the ROI of AI coding assistants and forecast expenses if moving to API-based pricing. Workflow Optimization by analyzing trace histories to identify repetitive, costly, or inefficient patterns in agent interactions. Post-Mortem Debugging of failed or erroneous agent sessions by using the full terminal replay and span-level inspection.

Unique Advantages

1. Differentiation: Unlike simple log parsers or local CLI tools (e.g., ccusage-style readers), LangWatch is a full-fledged, cloud-based (or self-hostable) OpenTelemetry platform. It aggregates data across all machines and persists history indefinitely, surviving local log cleanup. It offers far deeper analytics, visualization, and cross-session querying capabilities than basic terminal output.
2. Key Innovation: Its treatment of Cache Read and Cache Write as separate, priced token classes is a critical technical distinction. Many tracking solutions lump cache tokens into general input counts, which drastically misrepresents both total volume and cost, as cache reads can constitute the vast majority of tokens in long sessions. LangWatch's model-specific, class-aware pricing provides accurate theoretical cost calculations.
3. Key Innovation: The bidirectional integration via MCP is a significant architectural advantage. It doesn't just monitor the agent; it feeds insights back into the agent's context, enabling a form of meta-cognition where the AI can learn from and improve upon its own historical trace data.

Frequently Asked Questions (FAQ)

1. How does LangWatch track Claude Code usage without affecting performance? LangWatch uses the native OpenTelemetry export capability built into Claude Code. Running npx langwatch claude configures this export to send trace data asynchronously to the LangWatch backend. This has minimal performance overhead as it operates out-of-band from the agent's primary execution path.
2. Is my code or sensitive data secure with LangWatch? Yes. LangWatch employs automated redaction engines that scrub API secrets, keys, and PII from trace data before storage. For organizations with stricter requirements, self-hosted deployment options are available to keep all data within a private infrastructure.
3. Can I use LangWatch to track usage across my entire engineering team? Absolutely. LangWatch is designed for both individual developers and teams. It provides a centralized dashboard where managers can view aggregate cost, usage trends, and activity across all team members using supported AI coding agents, facilitating governance and budget management.
4. What is the difference between "Bundled" cost and "Theoretical Total" cost shown in the traces? "Bundled" indicates the session ran under a fixed-price subscription plan (e.g., Claude Max) where you are not directly billed per token. The "Theoretical Total" is the calculated cost if the same tokens (input, output, cache read/write at their respective rates) were consumed via the pay-as-you-go API, providing a benchmark for the value derived from the subscription.
5. Does LangWatch support AI coding agents other than Claude Code? Yes. The platform supports any agent that can export OpenTelemetry data to its OTLP endpoint. This includes Codex, OpenCode, Cursor, and Pi. All traces are unified in the same interface with consistent token accounting and cost analytics.