Honeycomb is built for questions you did not plan for — arbitrary dimensions, high cardinality, outliers rather than averages. Its MCP server puts that query model behind an agent, which suits the kind of open-ended digging that is tedious by hand.
https://mcp.honeycomb.io/mcp
Claude Sonnet 4.5
Chat with 60+ AI models on the same workflow — switch to a different model mid-conversation and re-run the same prompt, or use Compare mode to put several side-by-side and balance quality vs. cost.
A Honeycomb API key scoped to the environment you want readable. Keys are per-environment, so production and staging need separate ones.
How models use it and what it is built for.
The server exposes Honeycomb’s query surface over your datasets. An agent can list datasets and their columns to discover what dimensions exist, then compose a query with breakdowns, calculations, filters and orders — the same primitives as the query builder. That discovery step matters more here than in most observability tools, because Honeycomb’s value is in dimensions nobody thought to put on a dashboard, and a model that can enumerate columns will find them. Traces are retrievable for following a single slow request through services, and SLO and burn-alert state is readable so an agent can tie a spike to whether it is actually eating error budget. The workflow it enables is iterative: query, notice an outlier, break down by another dimension, repeat — which is exactly the loop that costs a human twenty minutes.
Typical tools an AI model can call. Exact names vary by version.
Copy any of these into MCP Agent Studio after connecting.
Which customer IDs are seeing p99 latency above two seconds on the API dataset?
Break down errors in the last hour by endpoint and then by build ID.
Fetch the slowest trace from the last thirty minutes and explain where the time went.
Are any SLOs burning error budget faster than usual today?
This is not a single-model product: you get the same MCP connection with 60+ models (Claude, GPT, Gemini, DeepSeek, open-weight, and more), you can switch mid-conversation, and you can open Compare mode to run the same prompt against multiple models at once. The card above is a suggested starting point for this server — not the only choice.
Default pick for Honeycomb
Claude Sonnet 4.5
Useful Honeycomb work is iterative refinement across several queries. Sonnet 4.5 keeps the earlier results in mind and narrows; cheaper models restart the investigation each turn.
Listing tools proves the server is reachable, not that a model can work with it. Evals go further: they read every tool on the server, write a test suite from its real schemas, and run it — code decides pass/fail on the responses (schema conformance, error codes, pagination, result caps) while a scoring model grades plain-English tasks driven through the tools.
Get a pass/fail report per tool with the evidence behind each verdict — and replay the same suite after every schema change. Destructive tools are excluded from the run.
Run evalsOpen MCP Agent Studio with the connection pre-filled. Add your token, pick any of 60+ models, and start chatting — no install required.
Open Agent StudioCommon questions about connecting, scoping and using it safely.
A hosted MCP server at mcp.honeycomb.io that exposes Honeycomb datasets, queries, traces and SLOs as tools, so an AI assistant can run the same high-cardinality investigations you would do in the query builder.
Because most outages are specific — one customer, one build, one region — and pre-aggregated dashboards average that away. Honeycomb keeps raw dimensions queryable, and an agent that can enumerate columns will try breakdowns a human would not have thought to check.
A Honeycomb API key for the environment you want to query. Keys are scoped per environment, so an agent given a staging key simply cannot read production — which makes this a clean place to start conservatively.
Yes. Honeycomb ingests OTLP, so if your services already emit OpenTelemetry traces the datasets are queryable through this server without any additional instrumentation work.
Read first. Alerting configuration is production configuration, and an agent that creates a noisy trigger costs your on-call rotation real sleep. Use it for investigation, and make alert changes a human decision.
Datadog
Query metrics, logs, traces and monitors from a conversation instead of a dashboard.
New Relic
Ask NRQL questions in plain English and get entity, alert and APM data back.
Grafana
Query Prometheus and Loki, search dashboards, triage alerts in chat.
ClickHouse
Run analytical SQL over billions of rows in natural language.