MCP Server for OpenTelemetry

Exposes OpenTelemetry trace data (spans, metrics, logs) through the Model Context Protocol (MCP) interface, allowing Claude and other MCP-compatible clients to query and analyze observability data without direct instrumentation. Implements MCP resource and tool handlers that translate OpenTelemetry SDK exports into structured JSON payloads compatible with LLM consumption, bridging observability backends (Jaeger, Datadog, etc.) with AI-driven analysis workflows.

Automatically enriches Claude's conversation context with relevant trace spans and metrics based on user queries about system behavior. Implements semantic matching between natural language questions (e.g., 'why is checkout slow?') and OpenTelemetry span attributes, then injects matched trace data into the prompt context. Uses MCP's context attachment mechanism to maintain trace lineage across multi-turn conversations.

Orchestrates multi-step root cause analysis by having Claude reason over traces, metrics, and logs to identify the underlying cause of issues. Implements a reasoning loop where Claude formulates hypotheses, requests specific traces or metrics to test them, and iteratively narrows down the root cause. Uses MCP's tool invocation pattern to enable Claude to request additional data as needed during analysis, without requiring upfront context injection.

Abstracts multiple OpenTelemetry exporters and trace backends (Jaeger, Datadog, Grafana Tempo, etc.) behind a unified MCP interface, normalizing span and metric schemas across different backend formats. Implements adapter pattern with backend-specific translators that convert proprietary trace formats into canonical OpenTelemetry JSON representation, allowing Claude to query traces from heterogeneous sources without backend-specific knowledge.

Exposes OpenTelemetry sampler configuration and span filtering rules as MCP tools, allowing Claude to dynamically adjust trace collection behavior based on analysis results. Implements MCP tool handlers that map to OpenTelemetry's Sampler interface, enabling Claude to request increased sampling for specific services or span attributes when investigating issues, without requiring application restarts.

Provides MCP tools for querying OpenTelemetry metrics (counters, histograms, gauges) with time-range and aggregation support, translating natural language metric queries from Claude into PromQL-like expressions. Implements metric backend abstraction that supports Prometheus, Grafana, and OpenTelemetry Metrics API, with built-in aggregation functions (sum, avg, percentile, rate) and time-series downsampling for efficient context injection.

Implements trace-to-log correlation by matching trace IDs and span IDs in log records with OpenTelemetry trace data, exposing correlated logs as MCP resources. Uses log backend APIs (ELK, Loki, Datadog) to retrieve logs with trace context, then enriches them with span metadata for unified analysis. Enables Claude to request logs for a specific trace and receive them pre-correlated without manual trace ID copying.

Introspects OpenTelemetry span attributes across collected traces to build a dynamic schema of available attributes, span types, and semantic conventions. Exposes this schema as MCP resources, allowing Claude to discover what span attributes are available and validate queries against the schema before execution. Implements schema caching with periodic updates to track schema evolution as new span types are introduced.

Analyzes historical trace data to establish baseline performance metrics (latency percentiles, error rates) and detects deviations that indicate regressions. Implements statistical comparison of recent spans against historical baselines, exposing regression alerts as MCP resources that Claude can query. Uses time-series analysis to identify which services or operations have degraded performance, enabling Claude to correlate regressions with recent changes.

Reconstructs service dependency graphs from trace data by analyzing span parent-child relationships and service names, exposing the dependency map as MCP resources. Generates visual representations (ASCII or JSON) of trace trees showing request flow across services, latency at each hop, and error propagation. Enables Claude to understand system architecture from traces and identify bottlenecks in request paths.

Applies statistical and machine learning techniques to identify unusual patterns in trace data, such as unexpected error rates, latency spikes, or unusual span sequences. Implements anomaly detection algorithms (isolation forest, z-score analysis) that learn normal trace behavior and flag deviations as MCP resources. Enables Claude to ask 'what's unusual about this trace?' and receive anomaly explanations without manual threshold configuration.