Writing Team
.png)
Most conversations about MCP (Model Context Protocol) stay theoretical. This session skipped the theory and went straight to a working list — nine specific MCP servers, what each one does, and real numbers on what changes when you plug it in. The presenter's own framing at the top: "I'll go over some of the MCP servers which will make your life easy if you are building agents or if you are building any tools, any workflow automations."
The Full List: Nine MCP Servers Worth Knowing
- Kubernetes MCP — AI-driven kubectl access across K8s, OpenShift, and K3s
- Grafana MCP — natural-language observability
- Terraform MCP — provider-aware infrastructure-as-code with drift detection
- ArgoCD MCP — GitOps managed through natural language
- PagerDuty MCP — 60+ tools for incident management
- GitLab MCP — 86+ tools covering code review, merge request operations, and pipelines
- n8n MCP — 1,396+ automation nodes, triggered by conversation
- FastMCP — a framework for building custom MCP servers in Python
- Security & Cost — role-based access control, prompt injection defense, and token budget management
Grafana MCP: Cutting Incident Triage From 28 Minutes to 8
This one had the clearest before-and-after numbers in the whole session. It bridges AI agents with Grafana for full observability control, exposing dashboards, data sources, alerts, and incidents as tools an AI can call directly, and enabling natural-language queries against Prometheus, Loki, and Tempo.
Without it: what the session called "dashboard tab-switching hell." Engineers manually opening four or five dashboards, writing PromQL by hand, cross-referencing logs, screenshotting panels for a ticket. Average investigation time: 28 minutes.
With it: incident response 3.5x faster, cutting triage from 28 minutes down to 8. It also finds unused metrics and suggests relabel configs to cut Prometheus cost, generating savings pull requests with zero PromQL expertise required. Repo: github.com/grafana/mcp-grafana, covering three MCP servers total (Core, Tempo tracing, and Loki logs).
PagerDuty MCP: Official Integration for AI-Driven Incident Management
This is PagerDuty's own first-party server, giving secure, controlled access to incident management from any MCP client, including Claude, VS Code, or Cursor. Write operations sit behind an --enable-write-tools flag specifically for safety.
It handles creating, acknowledging, and resolving incidents through natural language, querying incident history and timelines, managing on-call rosters and escalation policies, and triggering incident workflows with safety controls built in.
Without it: alert fatigue and slow triage, with context-switching wasting 10 to 15 minutes per incident. With it: a Cursor IDE /risk-score plugin that analyzes code changes against incident history before a deploy even happens, plus Azure Sentinel integration that auto-triages overnight alerts so incidents get summarized by 7am instead of waking someone up. Repo: github.com/PagerDuty/pagerduty-mcp-server.
GitLab MCP: Real Scenarios, Not Just Features
Two concrete workflows made this one land. First, an AI-driven code review loop: normally reviewer feedback sits in merge request comments while a developer context-switches between the GitLab UI and their editor. With MCP, the AI reads the comments, understands what's being asked, applies the fix, commits, and pushes without the developer leaving their editor.
Second, AI-powered MR descriptions: normally these are lazy ("fixed bug") or blank, forcing reviewers to read every line of the diff. With MCP, the AI reads the diff itself and generates a real description — what changed, why, which test cases are covered, and a walkthrough.
The session made a pointed case for MCP over the existing glab CLI: the CLI was built for humans, with help text parsing and guessed flags. MCP is AI-native — typed schemas let AI discover tools automatically, structured JSON removes parsing guesswork, and OAuth is built in. GitLab has since shipped glab mcp server directly.
n8n MCP: Automation by Conversation
This exposes n8n's workflow automation as MCP-callable tools, and it works in both directions — AI agents can trigger n8n workflows, and n8n workflows can call out to external MCP tools. It covers full CRUD on workflows, execution triggering and results retrieval, access to over 1,396 integration nodes (812 core plus 584 community), secure credential management, and AI-assisted error remediation.
Without it: building something like SOC automation means custom scripts, manual API integration management, and fragile webhook chains, locking non-developers out entirely. With it: the session described AI building a 7-node automation workflow through conversation, taking a Wazuh SOC integration from alert detection to endpoint isolation in seconds. Repo: github.com/czlonkowski/n8n-mcp.
FastMCP: Build Your Own in Python, 5x Faster
For anything not covered by an existing server, FastMCP is a Pythonic framework for building custom MCP servers, with a @mcp.tool() decorator for instant tool registration, automatic schema generation straight from Python type hints, a built-in browser debugger, and OpenAPI/FastAPI spec auto-conversion. The session's claim: 5x faster development compared to using the raw MCP SDK directly.
What To Do With This List
If your team is building agents or workflow automation right now, this list is close to a starter kit. Kubernetes, Grafana, Terraform, ArgoCD, PagerDuty, and GitLab cover most of a typical engineering and ops stack, n8n covers general automation, and FastMCP fills in whatever gap is left. The security and cost layer, RBAC, prompt injection defense, and token budgets, isn't optional once any of these are wired into a real production environment.
Wondering which of these fits into your own ops stack? Winsome helps companies figure out where automation actually belongs before wiring it in. Talk to Winsome about your AI tooling strategy.