Query and analyze metrics from a Prometheus server (PromQL, ranges, labels). Useful for observability and SRE workflows from inside an agent.
1 field
Credentials are injected into an ephemeral sandbox at boot and destroyed with it.
Light — starts in seconds
Sessions run up to 60 minutes before the sandbox is reclaimed.
60+ AI models
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.
How models use it and what it is built for.
The Prometheus MCP server lets a model run PromQL against your metrics — instant queries, range queries over a window, and label exploration to find out what is even being collected.
PromQL is the real barrier here. Most engineers can describe what they want ("p99 latency for the checkout service over the last hour, by pod") and cannot write the query from memory. A model that can both compose PromQL and immediately see whether the result looks sane closes that gap.
It needs only the server URL — no token — because Prometheus itself has no built-in auth. That makes setup trivial and makes network reachability the thing you actually have to think about.
Concrete operations exposed as tools.
In order. Each step assumes the previous one worked.
The sandbox reaches Prometheus over the public internet. A localhost or private-IP address will not work — use a tunnel (ngrok, Tailscale Funnel) or a reverse proxy with a public hostname.
Prometheus has no native auth, so anything publicly reachable is publicly readable. Front it with basic auth or an IP allowlist at the proxy before you expose it, even briefly.
The root URL of the server — https://prometheus.example.com — not the /graph or /api/v1 path.
Ask the model which metric names exist. A real list confirms the URL is right and the server is reachable.
What you provide before the sandbox starts.
Prometheus Server URLRequirede.g. https://prometheus.example.com or http://prom.internal:9090. Must be reachable from the public internet — localhost / private IPs need a tunnel (ngrok / Tailscale Funnel).
Paste any of these into Agent Studio once the server is connected.
What metrics are available for the checkout service? List the label values.
Show me p99 request latency over the last hour, broken down by pod.
Which containers had the highest memory usage in the last 24 hours?
Error rate spiked around 14:00 — query the relevant metrics and explain what happened.
Known constraints, stated plainly.
Deploy the server hosted, then watch which tools a model actually reaches for — with full JSON input and output on every call. Switch models mid-conversation to compare how each one uses the same server.
About the Prometheus MCP server.
No. Prometheus has no native authentication, so the server needs only a URL. That is exactly why you should put basic auth or an IP allowlist in front of any publicly reachable instance.
Only through a tunnel. Tailscale Funnel or ngrok will give the sandbox a reachable hostname. There is no way for an ephemeral cloud sandbox to reach a private VPC address directly.
No. This is a query surface over the read API — there is no ingestion or admin path exposed.
Use Prometheus directly when you want raw PromQL and fast iteration on a query. Use Grafana when the context you need is which dashboards and alerts exist, since those live in Grafana rather than in Prometheus.