OpenTelemetry MCP Server

Query and analyze LLM traces with AI assistance. Ask Claude to find expensive API calls, debug errors, compare model performance, or track token usage—all from your IDE.

An MCP (Model Context Protocol) server that connects AI assistants to OpenTelemetry trace backends (Jaeger, Tempo, Traceloop), with specialized support for LLM observability through OpenLLMetry semantic conventions.

See it in action:

https://github.com/user-attachments/assets/e2106ef9-0a58-4ba0-8b2b-e114c0b8b4b9

---

Table of Contents

• Quick Start
• Installation
• Features
• Configuration
• Tools Reference
• Example Queries
• Common Workflows
• Troubleshooting
• Development
• Support

---

Quick Start

No installation required! Configure your client to run the server directly from PyPI:

// Add to claude_desktop_config.json:
{
  "mcpServers": {
    "opentelemetry-mcp": {
      "command": "pipx",
      "args": ["run", "opentelemetry-mcp"],
      "env": {
        "BACKEND_TYPE": "jaeger",
        "BACKEND_URL": "http://localhost:16686"
      }
    }
  }
}

Or use uvx (alternative):

{
  "mcpServers": {
    "opentelemetry-mcp": {
      "command": "uvx",
      "args": ["opentelemetry-mcp"],
      "env": {
        "BACKEND_TYPE": "jaeger",
        "BACKEND_URL": "http://localhost:16686"
      }
    }
  }
}

That’s it! Ask Claude: “Show me traces with errors from the last hour”

---

Installation

For End Users (Recommended)

# Run without installing (recommended)
pipx run opentelemetry-mcp --backend jaeger --url http://localhost:16686

# Or with uvx
uvx opentelemetry-mcp --backend jaeger --url http://localhost:16686

This approach:

• ✅ Always uses the latest version
• ✅ No global installation needed
• ✅ Isolated environment automatically
• ✅ Works on all platforms

Per Client Integration

Prerequisites:

• Python 3.11 or higher
• pipx or uv installed

Features

Core Capabilities

• 🔌 Multiple Backend Support - Connect to Jaeger, Grafana Tempo, or Traceloop
• 🤖 LLM-First Design - Specialized tools for analyzing AI application traces
• 🔍 Advanced Filtering - Generic filter system with powerful operators
• 📊 Token Analytics - Track and aggregate LLM token usage across models and services
• ⚡ Fast & Type-Safe - Built with async Python and Pydantic validation

Tools

Backend Support Matrix

* Jaeger requires service_name parameter for span search

For Developers

If you’re contributing to the project or want to make local modifications:

# Clone the repository
git clone https://github.com/traceloop/opentelemetry-mcp-server.git
cd opentelemetry-mcp-server

# Install dependencies with UV
uv sync

# Or install in development mode with editable install
uv pip install -e ".[dev]"

---

Configuration

Supported Backends

Quick Configuration

Option 1: Environment Variables (Create .env file - see .env.example)

BACKEND_TYPE=jaeger
BACKEND_URL=http://localhost:16686

Option 2: CLI Arguments (Override environment)

opentelemetry-mcp --backend jaeger --url http://localhost:16686
opentelemetry-mcp --backend traceloop --url https://api.traceloop.com --api-key YOUR_KEY

Configuration Precedence: CLI arguments > Environment variables > Defaults

---

Usage

Quick Start with start_locally.sh (Recommended)

The easiest way to run the server:

./start_locally.sh

This script handles all configuration and starts the server in stdio mode (perfect for Claude Desktop integration). To switch backends, simply edit the script and uncomment your preferred backend.

Manual Running

For advanced use cases or custom configurations, you can run the server manually.

stdio Transport (for Claude Desktop)

Start the MCP server with stdio transport for local/Claude Desktop integration:

# If installed with pipx/pip
opentelemetry-mcp

# If running from cloned repository with UV
uv run opentelemetry-mcp

# With backend override (pipx/pip)
opentelemetry-mcp --backend jaeger --url http://localhost:16686

# With backend override (UV)
uv run opentelemetry-mcp --backend jaeger --url http://localhost:16686

HTTP Transport (for Network Access)

Start the MCP server with HTTP/SSE transport for remote access:

# If installed with pipx/pip
opentelemetry-mcp --transport http

# If running from cloned repository with UV
uv run opentelemetry-mcp --transport http

# Specify custom host and port (pipx/pip)
opentelemetry-mcp --transport http --host 127.0.0.1 --port 9000

# With UV
uv run opentelemetry-mcp --transport http --host 127.0.0.1 --port 9000

The HTTP server will be accessible at http://localhost:8000/sse by default.

Transport Use Cases:

• stdio transport: Local use, Claude Desktop integration, single process
• HTTP transport: Remote access, multiple clients, network deployment, sample applications

Tools Reference

1. search_traces

Search for traces with flexible filtering:

{
  "service_name": "my-app",
  "start_time": "2024-01-01T00:00:00Z",
  "end_time": "2024-01-01T23:59:59Z",
  "gen_ai_system": "openai",
  "gen_ai_model": "gpt-4",
  "min_duration_ms": 1000,
  "has_error": false,
  "limit": 50
}

Parameters:

• service_name - Filter by service
• operation_name - Filter by operation
• start_time / end_time - ISO 8601 timestamps
• min_duration_ms / max_duration_ms - Duration filters
• gen_ai_system - LLM provider (openai, anthropic, etc.)
• gen_ai_model - Model name (gpt-4, claude-3-opus, etc.)
• has_error - Filter by error status
• tags - Custom tag filters
• limit - Max results (1-1000, default: 100)

Returns: List of trace summaries with token counts

2. get_trace

Get complete trace details including all spans and OpenLLMetry attributes:

{
  "trace_id": "abc123def456"
}

Returns: Full trace tree with:

• All spans with attributes
• Parsed OpenLLMetry data for LLM spans
• Token usage per span
• Error information

3. get_llm_usage

Get aggregated token usage metrics:

{
  "start_time": "2024-01-01T00:00:00Z",
  "end_time": "2024-01-01T23:59:59Z",
  "service_name": "my-app",
  "gen_ai_system": "openai",
  "limit": 1000
}

Returns: Aggregated metrics with:

• Total prompt/completion/total tokens
• Breakdown by model
• Breakdown by service
• Request counts

4. list_services

List all available services:

{}

Returns: List of service names

5. find_errors

Find traces with errors:

{
  "start_time": "2024-01-01T00:00:00Z",
  "service_name": "my-app",
  "limit": 50
}

Returns: Error traces with:

• Error messages and types
• Stack traces (truncated)
• LLM-specific error info
• Error span details

Example Queries

Find Expensive OpenAI Operations

Natural Language: “Show me OpenAI traces from the last hour that took longer than 5 seconds”

Tool Call: search_traces

{
  "service_name": "my-app",
  "gen_ai_system": "openai",
  "min_duration_ms": 5000,
  "start_time": "2024-01-15T10:00:00Z",
  "limit": 20
}

Response:

{
  "traces": [
    {
      "trace_id": "abc123...",
      "service_name": "my-app",
      "duration_ms": 8250,
      "total_tokens": 4523,
      "gen_ai_system": "openai",
      "gen_ai_model": "gpt-4"
    }
  ],
  "count": 1
}

---

Analyze Token Usage by Model

Natural Language: “How many tokens did we use for each model today?”

Tool Call: get_llm_usage

{
  "start_time": "2024-01-15T00:00:00Z",
  "end_time": "2024-01-15T23:59:59Z",
  "service_name": "my-app"
}

Response:

{
  "summary": {
    "total_tokens": 125430,
    "prompt_tokens": 82140,
    "completion_tokens": 43290,
    "request_count": 487
  },
  "by_model": {
    "gpt-4": {
      "total_tokens": 85200,
      "request_count": 156
    },
    "gpt-3.5-turbo": {
      "total_tokens": 40230,
      "request_count": 331
    }
  }
}

---

Find Traces with Errors

Natural Language: “Show me all errors from the last hour”

Tool Call: find_errors

{
  "start_time": "2024-01-15T14:00:00Z",
  "service_name": "my-app",
  "limit": 10
}

Response:

{
  "errors": [
    {
      "trace_id": "def456...",
      "service_name": "my-app",
      "error_message": "RateLimitError: Too many requests",
      "error_type": "openai.error.RateLimitError",
      "timestamp": "2024-01-15T14:23:15Z"
    }
  ],
  "count": 1
}

---

Compare Model Performance

Natural Language: “What’s the performance difference between GPT-4 and Claude?”

Tool Call 1: get_llm_model_stats for gpt-4

{
  "model_name": "gpt-4",
  "start_time": "2024-01-15T00:00:00Z"
}

Tool Call 2: get_llm_model_stats for claude-3-opus

{
  "model_name": "claude-3-opus-20240229",
  "start_time": "2024-01-15T00:00:00Z"
}

---

Investigate High Token Usage

Natural Language: “Which requests used the most tokens today?”

Tool Call: get_llm_expensive_traces

{
  "limit": 10,
  "start_time": "2024-01-15T00:00:00Z",
  "min_tokens": 5000
}

---

Common Workflows

Cost Optimization

1. Identify expensive operations:

   Use get_llm_expensive_traces to find high-token requests
   

2. Analyze by model:

   Use get_llm_usage to see which models are costing the most
   

3. Investigate specific traces:

   Use get_trace with the trace_id to see exact prompts/responses
   

Performance Debugging

1. Find slow operations:

   Use get_llm_slow_traces to identify latency issues
   

2. Check for errors:

   Use find_errors to see failure patterns
   

3. Analyze finish reasons:

   Use get_llm_model_stats to see if responses are being truncated
   

Model Adoption Tracking

1. Discover models in use:

   Use list_llm_models to see all models being called
   

2. Compare model statistics:

   Use get_llm_model_stats for each model to compare performance
   

3. Identify shadow AI:

   Look for unexpected models or services in list_llm_models results
   

---

Development

Running Tests

# With UV
uv run pytest

# With coverage
uv run pytest --cov=openllmetry_mcp --cov-report=html

# With pip
pytest

Code Quality

# Format code
uv run ruff format .

# Lint
uv run ruff check .

# Type checking
uv run mypy src/

Troubleshooting

Backend Connection Issues

# Test backend connectivity
curl http://localhost:16686/api/services  # Jaeger
curl http://localhost:3200/api/search/tags  # Tempo

Authentication Errors

Make sure your API key is set correctly:

export BACKEND_API_KEY=your_key_here
# Or use --api-key CLI flag
opentelemetry-mcp --api-key your_key_here

No Traces Found

• Check time range (use recent timestamps)
• Verify service names with list_services
• Check backend has traces: curl http://localhost:16686/api/services
• Try searching without filters first

Token Usage Shows Zero

• Ensure your traces have OpenLLMetry instrumentation
• Check that gen_ai.usage.* attributes exist in spans
• Verify with get_trace to see raw span attributes

Future Enhancements

• [ ] Cost calculation with built-in pricing tables
• [ ] Model performance comparison tools
• [ ] Prompt pattern analysis
• [ ] MCP resources for common queries
• [ ] Caching layer for frequent queries
• [ ] Support for additional backends (SigNoz, ClickHouse)

Contributing

Contributions are welcome! Please ensure:

1. All tests pass: pytest
2. Code is formatted: ruff format .
3. No linting errors: ruff check .
4. Type checking passes: mypy src/

License

Apache 2.0 License - see LICENSE file for details

Related Projects

• OpenLLMetry - OpenTelemetry instrumentation for LLMs
• Model Context Protocol - MCP specification
• Claude Desktop - AI assistant with MCP support

Support

For issues and questions:

• GitHub Issues: https://github.com/traceloop/opentelemetry-mcp-server/issues
• PyPI Package: https://pypi.org/project/opentelemetry-mcp/
• Traceloop Community: https://traceloop.com/slack