read-honeycomb-opentelemetry

name	read-honeycomb-opentelemetry
description	Query Honeycomb for SLO compliance, error budgets, burn alerts, and OpenTelemetry traces. Use when asked about SLO status, service health, error budgets, latency, or Honeycomb data.
allowed-tools	mcp__honeycomb__get_slos, mcp__honeycomb__get_triggers, mcp__honeycomb__run_query, mcp__honeycomb__run_bubbleup, mcp__honeycomb__get_trace, mcp__honeycomb__find_columns, mcp__honeycomb__get_workspace_context, mcp__honeycomb__get_environment, mcp__honeycomb__get_dataset

Read Honeycomb OpenTelemetry

READ-ONLY observability skill. Query SLO status, error budgets, burn alerts, and traces.

📖 Full guide: Remote's SLOs framework (Notion)

Setup

If no Honeycomb MCP is configured:

claude mcp add honeycomb --transport http https://mcp.eu1.honeycomb.io/mcp

Only READ access is needed during OAuth setup.

Bootstrap

Always start by discovering what's available:

1. get_workspace_context → list environments and datasets.
2. get_environment → confirm the target environment (typically production).
3. get_dataset → confirm the target dataset.

Skip 2–3 if the user specifies environment and dataset explicitly.

SLO Status

Use get_slos with environment_slug to list all SLOs.

Triage pattern:

1. List all SLOs → identify any with compliance below target or triggered status.
2. For triggered SLOs, note remaining error budget and burn rate.
3. Use run_query to investigate contributing errors or latency.

Burn Alerts

Use get_triggers with environment_slug to list burn alert triggers and current state.

• Check which triggers are currently firing — these indicate active budget burn.
• Cross-reference fired triggers with SLO names to understand which SLOs are affected.

Diagnostic Queries

Use run_query with environment_slug and dataset_slug.

Error rates: COUNT where error = true or status_code >= 500, broken down by service.name or endpoint. Use RATE_AVG or COUNT with error filter vs total for error percentage.

Latency: HEATMAP(duration_ms) or P99(duration_ms) / P95(duration_ms) / P50(duration_ms). Break down by endpoint or operation.

Availability: Total requests vs error requests over a time window.

Time windowing: For incidents, narrow to the incident window first, then expand if needed.

BubbleUp Analysis

Use run_bubbleup to identify what makes a problematic subset different from the baseline.

When to use: SLO breaching, post-deploy comparison, latency spikes.

Pattern:

1. Define baseline (e.g., all requests in the last hour).
2. Define selection (e.g., requests where duration_ms > P99 or error = true).
3. BubbleUp surfaces disproportionate attributes (region, endpoint, deploy SHA).

Trace Investigation

MCP responses include URLs (trace_link from get_trace, Query Url from run_query). Always prefer those over manually constructing URLs.

Known Gotchas

15-second frontend timeout: Dragon cuts requests at 15s. Traces for long requests may appear incomplete, child spans disconnect from the parent. Gaps in span hierarchy on 15s+ traces are likely timeout-related, not missing instrumentation.

Constraints

• Read-only. Do not create, modify, or delete SLOs, triggers, or datasets.
• Always share Honeycomb URLs from query results and trace lookups in your response.
• Defaults. Use production environment unless told otherwise. Discover the dataset via bootstrap.