OpenTelemetry MCP Server: Give AI Agents Access to Traces, Metrics, and Distributed Debugging

Why Give AI Access to OpenTelemetry?

OpenTelemetry collects traces, metrics, and logs from your distributed systems. When requests slow down, error rates spike, or latency cascades across services, debugging means navigating your observability backend (Jaeger, Tempo, Honeycomb), filtering through thousands of spans, and manually correlating across service boundaries.

An MCP server for OpenTelemetry lets your AI agent search traces, query service metrics, find slow spans, and trace request paths directly from Claude, Cursor, or Windsurf through DataFaucet.

What Becomes Possible

Once your AI agent has OpenTelemetry access:

• "What's the P99 latency for the checkout service in the last hour?"
• "Find traces where the payment-service took over 2 seconds"
• "Which service in the request path is causing the latency spike?"
• "Show me error spans from the auth-service in the last 15 minutes"
• "What's the throughput of the order-processing service today vs yesterday?"
• "Trace the full request path for this failed order ID"

How It Works

1. Go to DataFaucet and enter your observability UI URL (Jaeger, Grafana Tempo, Honeycomb)
2. Authenticate with your observability platform credentials
3. Search traces, query metrics, filter spans as you normally would
4. Each API call becomes an MCP tool your AI can invoke
5. Deploy and connect to Claude, Cursor, or Windsurf

The AI agent accesses whatever trace and metric data your credentials allow.

Security Best Practices

• Read-only access — use viewer roles in your observability platform
• No trace deletion — prevent AI from removing observability data
• No alerting changes — block alert rule modifications through AI
• Service-scoped filters — restrict to specific service names if needed
• No PII in spans — ensure your instrumentation redacts sensitive attributes before AI access
• Time-bounded queries — limit lookback window to prevent expensive historical queries

Use Cases by Role

Backend developers: Performance debugging. "Why is this endpoint slow? Show me the trace breakdown" while staying in the editor writing the fix.

SRE/DevOps: Incident response. "Which service introduced latency? What's the error rate by service? Show me the dependency graph for failing requests."

Platform engineers: Observability review. "Which services have the highest span cardinality? Any services not instrumented? What's our trace sampling rate?"

Tech leads: Architecture understanding. "What services does the checkout flow depend on? What's the average hop count? Where are the bottlenecks?"

Observability UI vs AI Agent

Getting Started

Set up read access for your AI agent:

1. Configure viewer credentials for your observability backend:

- Jaeger: Read-only API access at /api/traces and /api/services

- Grafana Tempo: Viewer role in Grafana org

- Honeycomb: Read-only API key scoped to specific datasets

1. Point DataFaucet at your observability UI
2. Authenticate with the read-only credentials
3. Browse traces and metrics as you normally would