Claude Code OpenTelemetry Monitoring

Monitor your Claude Code usage with a complete observability stack: OpenTelemetry Collector, Prometheus, Loki, and Grafana. Track costs, token usage, cache efficiency, active time, and code changes in real-time with pre-built dashboards. Watch and check back for updates to this repo for step by step instructions for setting this up. Still fine-tuning the charts and panels.

An accompanying Claude Code usage metrics MCP server allows Claude Code access to it's own usage metrics (see below example along with example screenshots).

Based on Official Claude Code Documentation: Claude Code supports OpenTelemtry as described in Anthropic's official monitoring documentation.

Grafana Dashboards

Claude Code Usage Metrics MCP Server

claude mcp add --transport stdio metrics -s user -- uv run --directory /path/to/your/mcp-server metrics-server

claude mcp list
Checking MCP server health...

context7: https://mcp.context7.com/sse (SSE) - ✓ Connected
cf-docs: https://docs.mcp.cloudflare.com/sse (SSE) - ✓ Connected
metrics: uv run --directory /path/to/your/mcp-server metrics-server - ✓ Connected

Claude Code Usage Metrics MCP Server Demo

This section showcases the actual JSON responses from all 14 MCP tools, demonstrating the format and data you can expect.

1. get_current_cost

Returns today's total USD cost from Prometheus.

{
  "metric": "Total Cost Today",
  "value": 27.149809833783127,
  "formatted": "$27.1498",
  "unit": "currencyUSD"
}

2. get_token_usage

Returns token breakdown by type (input, output, cacheCreation, cacheRead).

{
  "metric": "Token Usage (24h)",
  "breakdown": {
    "cacheCreation": 2487212.387455661,
    "cacheRead": 40278591.98958367,
    "input": 16407.023220696385,
    "output": 283598.58088185877
  },
  "total": 43065809.98114189
}

3. get_cache_efficiency

Returns cache hit rate percentage.

{
  "metric": "Cache Hit Rate",
  "value": 93.52789643057625,
  "formatted": "93.53%",
  "unit": "percent"
}

4. get_available_metrics

Returns comprehensive reference of all Prometheus metrics and Loki log events (truncated for brevity).

{
  "prometheus_metrics": {
    "claude_code_cost_usage_USD_total": {
      "description": "Session cost in USD",
      "type": "counter",
      "labels": ["session_id", "model", "deployment_environment", "service_name"],
      "example_query": "sum by (session_id) (increase(claude_code_cost_usage_USD_total[24h]))"
    },
    "claude_code_token_usage_tokens_total": {
      "description": "Token usage by type",
      "type": "counter",
      "labels": ["type (input|output|cacheCreation|cacheRead)", "session_id", "model"],
      "example_query": "sum by (type) (increase(claude_code_token_usage_tokens_total[24h]))"
    }
  },
  "loki_log_events": {
    "tool_result": {
      "event_name_label": "tool_result",
      "description": "Tool execution results",
      "attributes": ["tool_name", "duration_ms (numeric)", "success (true|false)", "error", "decision_source", "session_id"],
      "example_query": "{service_name=\"claude-code\", event_name=\"tool_result\"} | unwrap duration_ms",
      "note": "Use label selectors, NOT | json filter (OTLP format)"
    },
    "user_prompt": {
      "event_name_label": "user_prompt",
      "description": "User prompt submissions (privacy-sensitive)",
      "attributes": ["prompt (full text)", "prompt_length (numeric)", "session_id", "model"],
      "example_query": "{service_name=\"claude-code\", event_name=\"user_prompt\"}",
      "note": "Requires OTEL_LOG_USER_PROMPTS=1 environment variable"
    }
  },
  "common_patterns": {
    "promql_patterns": [
      "Cache efficiency: sum(increase(...{type=\"cacheRead\"}[24h])) / clamp_min(sum(increase(...[24h])), 1) * 100",
      "Per-session cost: sum by (session_id) (increase(claude_code_cost_usage_USD_total[24h]))"
    ],
    "logql_patterns": [
      "Slow tools: {service_name=\"claude-code\", event_name=\"tool_result\"} | duration_ms > 5000",
      "Tool failures: {service_name=\"claude-code\", event_name=\"tool_result\", success=\"false\"}"
    ]
  }
}

5. get_recent_prompts

Returns recent user prompt submissions (requires OTEL_LOG_USER_PROMPTS=1).

{
  "count": 10,
  "prompts": [
    {
      "prompt": "metrics MCP server is installed @mcp-server/README.md at claude code user scope, I need for you to test each mcp tool to ensure it's working",
      "length": "140",
      "timestamp": "2025-10-23T18:23:01.865Z",
      "model": "N/A",
      "session_id": "dcc1d3a8-fded-48db-827b-fcf1c1f3a9d3"
    },
    {
      "prompt": "what is my token usage breakdown",
      "length": "32",
      "timestamp": "2025-10-23T17:56:27.830Z",
      "model": "N/A",
      "session_id": "fd9afd28-606a-4e78-b115-5ad5443082fb"
    }
  ],
  "truncated": false
}

6. get_tool_results

Returns recent tool execution logs with performance metrics (truncated for brevity).

{
  "count": 20,
  "results": [
    {
      "tool_name": "mcp__metrics__get_cache_efficiency",
      "duration_ms": "26",
      "success": "true",
      "error": "",
      "decision_source": "user_permanent",
      "timestamp": "2025-10-23T18:24:23.528Z",
      "session_id": "dcc1d3a8-fded-48db-827b-fcf1c1f3a9d3"
    },
    {
      "tool_name": "Bash",
      "duration_ms": "1354",
      "success": "true",
      "error": "",
      "decision_source": "config",
      "timestamp": "2025-10-23T18:20:40.931Z",
      "session_id": "e24ee0d8-4c7a-416e-aa8e-767fbbadb75c"
    }
  ],
  "truncated": false
}

7. query_prometheus

Execute arbitrary PromQL queries against Prometheus.

Example query: sum(increase(claude_code_cost_usage_USD_total[24h]))

{
  "query": "sum(increase(claude_code_cost_usage_USD_total[24h]))",
  "value": 27.303288749445116,
  "raw_result": {
    "status": "success",
    "data": {
      "resultType": "vector",
      "result": [
        {
          "metric": {},
          "value": [1761243919.881, "27.303288749445116"]
        }
      ]
    }
  }
}

8. query_loki

Execute arbitrary LogQL queries against Loki.

Example query: {service_name="claude-code"} |= "tool_result" (limit 5)

{
  "query": "{service_name=\"claude-code\"} |= \"tool_result\"",
  "count": 5,
  "log_lines": [
    "claude_code.tool_result",
    "claude_code.tool_result",
    "claude_code.tool_result",
    "claude_code.tool_result",
    "claude_code.tool_result"
  ],
  "truncated": false
}

Dashboard Analysis Tools

9. list_dashboard_panels

Lists all panels from both Grafana dashboards (truncated for brevity - shows 103 total panels).

{
  "total_panels": 103,
  "panels": [
    {
      "id": 28,
      "title": "Real-Time Cost Burn Rate ($/hour)",
      "type": "stat",
      "description": "Current cost burn rate in dollars per hour based on last 5 minutes of activity",
      "has_queries": true
    },
    {
      "id": 27,
      "title": "Total Cost Today",
      "type": "stat",
      "description": "Total Cost Today",
      "has_queries": true
    },
    {
      "id": 3,
      "title": "Cache Hit Rate %",
      "type": "gauge",
      "description": "Cache Hit Rate %",
      "has_queries": true
    }
  ]
}

10. find_panel_by_name

Search panels by keyword (case-insensitive, partial match).

Example search: "cost"

{
  "search": "cost",
  "matches": 42,
  "results": [
    {
      "id": 28,
      "title": "Real-Time Cost Burn Rate ($/hour)",
      "type": "stat",
      "description": "Current cost burn rate in dollars per hour based on last 5 minutes of activity"
    },
    {
      "id": 27,
      "title": "Total Cost Today",
      "type": "stat",
      "description": "Total Cost Today"
    },
    {
      "id": 36,
      "title": "Cost Forecast",
      "type": "stat",
      "description": "Projected daily and monthly cost based on current usage patterns"
    }
  ]
}

11. get_panel_query

Extract exact PromQL/LogQL queries from a specific panel.

Example panel: "Total Cost Today"

{
  "panel": "Total Cost Today",
  "queries": [
    "sum(increase(claude_code_cost_usage_USD_total[24h]))",
    "avg_over_time((sum(increase(claude_code_cost_usage_USD_total[24h])))[7d:24h])"
  ],
  "primary_query": "sum(increase(claude_code_cost_usage_USD_total[24h]))"
}

12. explain_panel_query

Explain a panel's query with breakdown and live value.

Example panel: "Cache Hit Rate"

{
  "panel": "Cache Hit Rate %",
  "description": "Cache Hit Rate %",
  "query": "sum(increase(claude_code_token_usage_tokens_total{type=\"cacheRead\"}[24h])) /\nclamp_min(sum(increase(claude_code_token_usage_tokens_total[24h])), 1) * 100",
  "explanation": {
    "query": "sum(increase(claude_code_token_usage_tokens_total{type=\"cacheRead\"}[24h])) /\nclamp_min(sum(increase(claude_code_token_usage_tokens_total[24h])), 1) * 100",
    "components": [
      "sum - Total across all series",
      "increase - Total change over time range",
      "clamp_min - Prevent division by zero",
      "[24h] - Over the last 24h",
      "* 100 - Multiply by 100",
      "Filter: type=\"cacheRead\""
    ],
    "meaning": "This query calculates metrics from claude_code_token_usage_tokens_total. sum - Total across all series increase - Total change over time range clamp_min - Prevent division by zero"
  },
  "current_value": 93.53188202200037,
  "unit": null
}

13. explain_promql_query

Break down the structure of a PromQL query.

Example query: `rate(claude_code_to