How a DevOps Team Connected PagerDuty, Grafana, and GitHub Actions to AI

The Problem: Incident Triage Was All Context Gathering

A 3-person DevOps team at a Series B startup ran 40+ microservices. On-call rotation meant one person fielding every alert. Each page required: check PagerDuty for alert details, open Grafana for metrics context, then dig through GitHub Actions to see what deployed recently.

By the time they had enough context to decide if the alert was real, 20 minutes had passed. Half the time it was a false positive from a config change nobody mentioned in Slack.

What They Connected

Three MCP servers, created in under a minute each with DataFaucet:

1. PagerDuty (incidents, services, on-call schedules)

Browsed the PagerDuty dashboard. Alert details, incident history, service dependencies, and escalation policies became queryable through AI.

2. Grafana (dashboards, metrics queries, alert rules)

Browsed Grafana panels for 60 seconds. CPU, memory, error rates, and latency metrics became accessible. The AI could check the same dashboards the team would check manually.

3. GitHub Actions (workflow runs, deploy history, failure logs)

Browsed the Actions tab on their main repos. Recent deploys, failed workflows, and commit history became tool calls.

The Workflow Now

Alert fires at 2am. On-call engineer opens Claude:

"PagerDuty alert PD-4521 just fired. What service is affected, what do the Grafana metrics look like for the last hour, and did anything deploy in the last 2 hours?"

One prompt. Three tool calls. Response in 15 seconds:

• Alert is on the payments-api service (p95 latency spike)
• Grafana shows latency jumped from 120ms to 2.4s at 1:47am
• GitHub Actions shows a config change deployed at 1:45am to payments-api (PR #847: updated connection pool settings)

Root cause identified before the engineer fully wakes up. Rollback the deploy, acknowledge the incident, go back to sleep.

The Numbers

Why It Worked

The team didn't need AI to fix incidents. They needed AI to gather context fast enough to make good decisions. The three-tool combo covers the critical question: "What changed, what broke, and how bad is it?"

The connection pool config change in the example above took 2 minutes to roll back. The expensive part was always the 20 minutes before someone figured out what happened.

Setup Details

All three servers deployed as hosted SSE endpoints. Added to Cursor (for during-code debugging) and Claude Desktop (for on-call response):

{
  "mcpServers": {
    "pagerduty": {
      "url": "https://datafaucet.dev/api/mcp/SERVER_1/sse"
    },
    "grafana": {
      "url": "https://datafaucet.dev/api/mcp/SERVER_2/sse"
    },
    "github-actions": {
      "url": "https://datafaucet.dev/api/mcp/SERVER_3/sse"
    }
  }
}

Total setup time: 4 minutes. No custom code. No API docs needed (PagerDuty and Grafana APIs are well-documented but nobody wanted to spend a day writing MCP servers for each).

The on-call engineer's take: "I used to dread pages because the first 20 minutes were always the same boring context-gathering. Now I skip straight to the decision."

Create your own MCP servers for monitoring and DevOps tools at datafaucet.dev.