TL;DR
The Monitoring Stack is 5 MCPs (Sentry, Grafana, Postgres, Slack, GitHub) that collapse incident investigation into one Claude prompt. Correlating error spikes, latency, deploys, and DB state — which usually takes 5 tabs and 10 minutes — becomes 90 seconds. Essential for any on-call engineer.
The Monitoring & Observability Stack
Errors, metrics, alerting, and on-call — unified in one AI session
Install the whole stack
$ mcpizy install sentry grafana postgres slack githubOne command installs and configures all 5 MCPs for Claude Code, Cursor, Windsurf, or any MCP-compatible client.
Why this stack?
Observability is correlating signals across tools: 'error spike in Sentry' + 'latency spike in Grafana' + 'deploy in GitHub 5 min ago' + 'DB CPU high in Postgres'. Humans do this context-stitching manually during incidents, under pressure. MCPs collapse it: one Claude prompt pulls from all 5 sources and hands you the likely root cause.
This stack is PagerDuty-adjacent (not a replacement) — it's the investigation layer. The first 10 minutes of any incident usually lives here, and MCPs can compress those 10 minutes to 90 seconds.
MCPs in this stack (5)
What this stack lets you do
New Sentry error → triage in 60 seconds
See recipe- 1Sentry fires new error alert to Slack
- 2Claude fetches the full stack trace + affected users
- 3Correlates with recent GitHub commits via file paths
- 4Checks Grafana for matching latency/error rate spikes
- 5Proposes rollback or fix, drafts Linear ticket
- 6Posts status to #incidents
SLO breach investigation
See recipe- 1Grafana alert fires — SLO breached
- 2Claude pulls the exact breach window from Grafana
- 3Queries Postgres for slow queries in that window
- 4Checks Sentry for coincident errors
- 5Links to the deploy that likely introduced the regression
- 6Status summary posted to Slack
Weekly reliability report
- 1Claude summarises last week's Sentry error trends
- 2Grafana dashboard data pulled for SLO attainment
- 3GitHub queried for hotfix PRs and rollbacks
- 4Slack digest posted every Monday at 9am
- 5Insights logged to Notion for the retro
Estimated value
50% reduction in mean-time-to-mitigation during incidents. For a team handling 10 incidents/month, that's ~15 hours of engineer time reclaimed, plus the compounding effect of less burnout.
Frequently asked questions
Does this replace Datadog?
No — Datadog (or your existing APM) stays as the collection layer. MCPs are the query/investigation layer. If you use Datadog, add the Datadog MCP (community) in place of or alongside Grafana MCP.
Can Claude actually trigger a rollback?
Via GitHub MCP, yes — it can revert a commit or redeploy a previous tag. Most teams gate this behind a human confirmation for production. For non-prod environments, full automation is fine.
Is 5 MCPs enough for observability?
For most SaaS teams under ~50 engineers, yes. Large shops add PagerDuty MCP for rotation, Datadog MCP for APM, and possibly Honeycomb MCP for traces. The 5 above are the core.
What about logs?
Grafana Loki MCP covers log search. For CloudWatch Logs, AWS MCP handles it. Logs are the biggest gap in the 5-MCP minimum — if logs matter to you, add Loki or CloudWatch-focused MCP as MCP #6.
How much does this cost to run?
$0 at the MCP layer — all open source. You pay for the underlying tools (Sentry free for <5K errors/mo, Grafana Cloud free for 3 users, Slack free). Full monitoring stack often fits under $100/mo for a small team.