jaeger-mcp

<!-- mcp-name: io.github.mshegolev/jaeger-mcp -->

MCP server for [Jaeger](https://www.jaegertracing.io/) distributed tracing. Give Claude (or any MCP-capable agent) read access to your trace data — search traces, inspect spans, compare traces, compute span statistics, map service dependencies, predict performance issues, and forecast capacity needs — without leaving the conversation.

Why another Jaeger MCP?

The existing Jaeger integrations require a running UI or custom scripts. This server:

• Speaks the standard Model Context Protocol over stdio — works with Claude Desktop, Claude Code, Cursor, and any MCP client.
• Is read-only: all 12 tools carry readOnlyHint: true — zero risk of modifying trace data.
• Returns dual-channel output: structured JSON (structuredContent) for programmatic use + Markdown (content) for human-readable display.
• Has actionable error messages that name the exact env var to fix and suggest a next step.
• Supports Bearer token, HTTP Basic auth, or no auth (common for internal deployments).
• Includes OpenAPI specification documenting the underlying Jaeger Query API (openapi.yaml).

Tools

Installation

pip install jaeger-mcp

Or run directly without installing:

uvx jaeger-mcp

Configuration

All configuration is via environment variables:

Copy .env.example to .env and fill in your values.

Claude Desktop / Claude Code setup

Add to your MCP config (claude_desktop_config.json or .claude/mcp.json):

{
  "mcpServers": {
    "jaeger": {
      "command": "jaeger-mcp",
      "env": {
        "JAEGER_URL": "https://jaeger.example.com",
        "JAEGER_TOKEN": "your-token-here"
      }
    }
  }
}

Or with uvx (no install required):

{
  "mcpServers": {
    "jaeger": {
      "command": "uvx",
      "args": ["jaeger-mcp"],
      "env": {
        "JAEGER_URL": "https://jaeger.example.com"
      }
    }
  }
}

Docker

docker run --rm -e JAEGER_URL=https://jaeger.example.com jaeger-mcp

Example queries

Once configured, ask Claude:

• "What services does Jaeger know about?"
• "Find traces with HTTP 500 errors in order-service from the last hour"
• "Show me the slowest traces (over 2 seconds) for GET /checkout"
• "What caused the error in trace abcdef1234567890?"
• "Map the service dependency graph for the last 7 days"
• "Which services call postgres most frequently?"
• "Compare trace abc123 against trace def456 — what spans changed?"
• "What are the p95 latencies per operation in order-service?"

Tool usage guide

jaeger_list_services

Returns all service names Jaeger has seen. Start here when you don't know which services are instrumented. Output is capped at 500 services with a truncation hint.

jaeger_list_operations

Returns all operation names for a given service (e.g. HTTP route names, gRPC method names). Use to discover valid operation names before filtering jaeger_search_traces.

jaeger_search_traces

The main search tool. Filters:

• service (required) — service name from jaeger_list_services
• operation — narrow to a specific endpoint
• tags — JSON string of tag filters, e.g. {"http.status_code":"500"} or {"error":"true"}
• start / end — time range in microseconds UTC
• min_duration / max_duration — duration strings like "100ms", "1.5s", "2m"
• limit — default 20, max 1500

Returns trace summaries with trace_id, duration_us, span_count, service_count, root_operation, errors_count.

jaeger_get_trace

Full trace detail. Accepts a trace_id (hex string, 16-32 chars) and returns:

• All spans with tags, service names, parent/child relationships
• Per-service statistics (span count, total duration, error count)
• Execution tree (each node lists its child span IDs)

Error spans are identified by tags["error"] = "true".

jaeger_get_dependencies

Service topology graph. Returns directed edges (parent → child) with call_count. Use lookback_hours (default 24, max 720) to control the window.

jaeger_compare_traces

Structural diff between two traces. Accepts two trace_id hex strings and matches spans by (operationName, serviceName, parentOperation) — not span ID. Reports:

• Added spans — present in trace B but not trace A
• Removed spans — present in trace A but not trace B
• Changed spans — matched but differ in duration or tags (shows deltas)
• Unchanged count — number of identical spans

Use to compare a slow trace against a fast one, or to see what changed between deployments.

jaeger_span_statistics

Per-operation latency percentiles and error rates. Fetches up to limit traces (default 20, max 100) for a service and aggregates all spans by operation name. Reports per operation:

• count — total spans observed
• p50_duration_us, p95_duration_us, p99_duration_us — latency percentiles
• error_count, error_rate — errors (identified by tags["error"] = "true")

Use to find the slowest or most error-prone operations in a service.

jaeger_critical_path

Identifies the longest-duration span chain from root to leaf in a trace (the critical path) and ranks spans by self-time to find performance bottlenecks.

Reports:

• Critical path spans with operation, service, duration, and percentage-of-total
• Bottleneck spans ranked by exclusive duration (self-time)

Use to answer "Why is this trace so slow?" and "Which operations consume the most CPU/self-time?"

jaeger_compare_windows

Compares aggregate trace behavior between two time periods for a service to detect performance regressions or improvements across deployments.

Reports:

• Per-operation diff summary showing added, removed, faster, slower operations
• Deviation scoring with numeric scores per operation and overall
• Latency percentile changes (p50, p95) and error rate deltas

Use to answer "Did our latest deployment affect performance?" and "Which operations got slower after the database upgrade?"

jaeger_detect_anomalies

Scans for statistically significant latency spikes or error-rate increases in a service's recent traces compared to historical baselines.

Reports:

• Flagged operations with anomaly type (latency or error_rate)
• Severity classification (low to critical) with z-scores
• Current vs baseline values for affected metrics

Use to proactively identify performance degradations and reliability issues before they impact users.

Library facade (in-process use)

jaeger-mcp can also be used as a Python library without an MCP server:

from jaeger_mcp import JaegerClient

client = JaegerClient.from_env()  # reads JAEGER_URL from env
trace = client.get_trace("abcdef1234567890abcdef1234567890")

for span in trace.spans:
    if span.error:
        print(f"{span.service_name}: {span.operation} at {span.start_utc}")
        print(f"  tags: {span.tags}")

Available methods: get_trace(), search_traces(), list_services(), get_dependencies(), compare_traces(), span_statistics(), critical_path(), compare_windows(), detect_anomalies().

Domain objects: Span, Trace, TraceSummary, ServiceDep, TraceComparison, SpanIdentity, SpanChange, SpanStatisticsResult, OperationStatResult, CriticalPathOutput, CriticalPathSpan, BottleneckSpan, WindowComparisonOutput, OperationDiff, AnomalyDetectionOutput, OperationAnomaly — all with typed fields.

API Documentation

This project includes comprehensive OpenAPI specifications in the docs/ directory:

1. Jaeger Query Service API (openapi.yaml) - Documents the actual Jaeger API endpoints
2. MCP Tools API (docs/mcp-tools-openapi.yaml) - Documents the MCP tools as conceptual HTTP endpoints

These specifications are useful for:

• Understanding the underlying API calls made by each tool
• Developing alternative integrations
• Debugging API interactions
• Generating client libraries or documentation

See docs/README.md for more details on both specifications.

Performance characteristics

• All tools use a single persistent requests.Session with connection pooling.
• The session has trust_env = False to bypass environment proxies (Jaeger is typically an internal service).
• Requests time out after 30 seconds (configurable via JAEGER_TIMEOUT).
• Transient HTTP errors (429/5xx) are retried with exponential backoff (configurable via JAEGER_RETRY_ATTEMPTS).
• list_services and list_operations responses are cached for 120 seconds (configurable via JAEGER_CACHE_TTL).
• jaeger_search_traces passes limit directly to Jaeger — avoid requesting more traces than needed.
• jaeger_get_trace fetches the full trace in one call — large traces (thousands of spans) may be slow.
• jaeger_get_dependencies aggregates over the full lookback window; large windows may be slow on busy clusters.

Development

git clone https://github.com/mshegolev/jaeger-mcp
cd jaeger-mcp
pip install -e '.[dev]'
pytest tests/ -v
ruff check src tests
ruff format src tests

License

MIT — see LICENSE.