Skip to content

How to Monitor Claude Code with OpenTelemetry

Published: Tuesday, Feb 24, 2026 • 🏷️ claude code • otel • grafana • prometheus • observability • monitoring

thumbnail

TL;DR: Claude Code is powerful but you have minimal visibility into costs, token usage, or development patterns. I pre-configured a 5-minute OpenTelemetry monitoring stack (Prometheus + Loki + Grafana) that shows exactly where your money goes and how effective prompt caching really is.

The Problem

Claude Code gives you no built-in visibility into:

  • Which models are eating your budget
  • Whether prompt caching is actually saving you money
  • How many tokens you’re burning through daily
  • Development velocity metrics (lines changed, commits, PRs)

The Solution

I configured an observability stack using industry-standard tools:

pic1

Architecture:

  • Claude Code exports telemetry via OpenTelemetry Protocol (OTLP)
  • OTel Collector receives metrics and logs, routes them appropriately
  • Prometheus stores time-series metrics (costs, tokens, cache rates)
  • Loki stores logs (session logs, real-time activity)
  • Grafana visualizes everything with pre-built dashboards

All configuration, automation, and dashboards are in the GitHub repo: halyph/claude-mon.

Quick Demo

One command:

git clone https://github.com/halyph/claude-mon.git
cd claude-mon
./setup.sh

setup_sh

Wait 2 minutes for setup to complete, then open http://localhost:3000. Default credentials: admin/admin.

What You Get

Two Grafana Dashboards

Metrics Dashboard

pic2

16 panels:

  • Cost tracking: Total spend by model, cost distribution pie chart
  • Token usage: Input/output/cache tokens over time
  • Cache effectiveness: Hit rate gauge showing cache read/write ratios
  • Development activity: Lines changed, commits created, PRs opened
  • Code edits: Accept/reject decisions by tool and language

Logs Dashboard

pic3

Real-time log streaming with:

  • Live session activity (refreshes every 5 seconds)
  • Log volume over time
  • Per-session filtering
  • Model-based categorization

NOTE: Dashboards are pre-configured templates. Customize as needed.

Key Metrics Reference

Claude Code exports these metrics (converted to Prometheus format):

Metric What It Measures Why It Matters
claude_code_cost_usage_total Cost in USD by model/session Budget tracking, model comparison
claude_code_token_usage_total Tokens by type (input/output/cache) Optimize prompts, track cache effectiveness
claude_code_active_time_total Session duration (user/cli) Understand usage patterns
claude_code_lines_of_code_count Lines added/removed Development velocity
claude_code_commit_count Git commits created Productivity metric
claude_code_code_edit_tool_decision Accept/reject by tool/language Quality tracking

All metrics include standard labels: model, session_id, terminal_type, user_id, type.

Data Persistence

Docker volumes preserve all historical data:

  • prometheus-data: Metrics time-series
  • loki-data: Log streams
  • grafana-data: Dashboards and settings
docker-compose down      # Stops services, keeps data
docker-compose down -v   # Stops services, deletes data

References