PerformanceMonitor MCP — 사용 사례, 설치 & 라이브 데모

왜 쓰나요

핵심 기능

50+ diagnostic tools across CPU, memory, waits, blocking, deadlocks, file I/O, plans
Built-in 30-rule PlanAnalyzer for surfacing common query anti-patterns
Localhost-only HTTP binding — never exposed on the network
Works against SQL Server 2016-2025, Azure SQL, and AWS RDS
Read-only diagnostics — cannot modify data, schema, or server config

라이브 데모

실제 사용 모습

performancemonitor.replay ▶ 준비됨

0/0

설치

클라이언트 선택

~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "performancemonitor": {
      "command": "TODO",
      "args": [
        "See README: https://github.com/erikdarlingdata/PerformanceMonitor"
      ],
      "_inferred": true
    }
  }
}

Claude Desktop → Settings → Developer → Edit Config 열기. 저장 후 앱 재시작.

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "performancemonitor": {
      "command": "TODO",
      "args": [
        "See README: https://github.com/erikdarlingdata/PerformanceMonitor"
      ],
      "_inferred": true
    }
  }
}

Cursor는 Claude Desktop과 동일한 mcpServers 스키마 사용. 프로젝트 설정이 전역보다 우선.

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "performancemonitor": {
      "command": "TODO",
      "args": [
        "See README: https://github.com/erikdarlingdata/PerformanceMonitor"
      ],
      "_inferred": true
    }
  }
}

Cline 사이드바의 MCP Servers 아이콘 클릭 후 "Edit Configuration" 선택.

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "performancemonitor": {
      "command": "TODO",
      "args": [
        "See README: https://github.com/erikdarlingdata/PerformanceMonitor"
      ],
      "_inferred": true
    }
  }
}

Claude Desktop과 같은 형식. Windsurf 재시작 후 적용.

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "performancemonitor",
      "command": "TODO",
      "args": [
        "See README: https://github.com/erikdarlingdata/PerformanceMonitor"
      ]
    }
  ]
}

Continue는 맵이 아닌 서버 오브젝트 배열 사용.

~/.config/zed/settings.json

{
  "context_servers": {
    "performancemonitor": {
      "command": {
        "path": "TODO",
        "args": [
          "See README: https://github.com/erikdarlingdata/PerformanceMonitor"
        ]
      }
    }
  }
}

context_servers에 추가. 저장 시 Zed가 핫 리로드.

claude mcp add performancemonitor -- TODO 'See README: https://github.com/erikdarlingdata/PerformanceMonitor'

한 줄 명령. claude mcp list로 확인, claude mcp remove로 제거.

사용 사례

실전 활용법: PerformanceMonitor

Triage a slow SQL Server during an incident

👤 DBAs, SREs on call for database issues ⏱ ~5 min intermediate

언제 쓸까: Users are reporting timeouts and you need to know what's blocking whom in the next 60 seconds.

사전 조건

PerformanceMonitor app installed and connected to target server — Download from the GitHub releases and point it at your monitoring DB
Target login has VIEW SERVER STATE — GRANT VIEW SERVER STATE TO [monitor_login]
MCP registered with your client — claude mcp add --transport http --scope user sql-monitor http://localhost:5151/

흐름

Get the health overview

Run the health check — any red flags on waits, blocking, or memory pressure right now?✓ 복사됨

→ Short list of elevated metrics
Drill into the top offender

If blocking is the top issue, show me the head blocker chain and the SQL text of the blocking session.✓ 복사됨

→ Specific SPID + query text
Get a recommendation

Should I kill that SPID, wait it out, or is there a missing index that would prevent this? Be specific.✓ 복사됨

→ One clear recommended action with rationale

결과: A diagnosed incident with a specific next action — kill, wait, index, or escalate — in under 5 minutes.

함정

Killing a blocker that was about to commit can cause more damage — Always ask for how long it's been running and whether it holds an open transaction before killing
Metrics from the monitoring DB lag the real server by the collector interval — For live view, use the 'current state' tools rather than aggregated history

함께 쓰기: sentry

Analyze a slow query plan with AI-assisted review

👤 Backend devs and DBAs tuning queries ⏱ ~15 min intermediate

언제 쓸까: A query is slow and the plan is 200 operators — you don't want to read it manually.

흐름

Fetch the plan

Get the most recent plan for query hash 0xA1B2C3 and run the PlanAnalyzer against it.✓ 복사됨

→ Plan + list of rule hits
Explain the worst rule hit

Explain the #1 rule hit in plain English — what's happening and why it's bad.✓ 복사됨

→ 2-3 sentence explanation in English, not jargon
Suggest a fix

Suggest the smallest change to fix it: index, rewrite, or stats update. Estimate impact.✓ 복사됨

→ Concrete DDL or rewrite

결과: A specific tuning action backed by both plan evidence and the PlanAnalyzer rule.

함정

Plan was cached under a different parameter sniffing path — Regenerate the plan with OPTION(RECOMPILE) before trusting its shape

Produce a weekly SQL Server health summary

👤 Database owners who want passive monitoring, not live on-call ⏱ ~20 min beginner

언제 쓸까: Friday afternoon wrap-up; know before users tell you.

흐름

Pull week-over-week stats

Compare top waits and top queries this week vs last week. Flag anything that got materially worse.✓ 복사됨

→ Deltas with percentages
Draft a team-readable summary

Write a one-page health summary — good/bad/ugly — for the engineering channel.✓ 복사됨

→ Markdown post ready to paste

결과: A shareable weekly health post with concrete items to track next week.

함께 쓰기: sentry

조합

다른 MCP와 조합해 10배 효율

performancemonitor + sentry

Sentry flags a timeout spike; cross-reference with SQL Server blocking at the same timestamp

Sentry event at 14:03 UTC shows SqlException. What was happening on the DB between 14:02 and 14:04 — any blocking or waits spikes?✓ 복사됨

performancemonitor + github

Find the commit that introduced a slow query

Query hash 0xA1B2 started showing up in the top 10 on 2026-03-22. Find GitHub commits to the data-access layer around that date.✓ 복사됨

도구

이 MCP가 노출하는 것

도구	입력	언제 호출	비용
get_top_waits	server_name?: str, hours_back?: int	First stop in any triage — what is the server actually waiting on	1 SQL query on the monitoring DB
get_active_blocking	server_name?: str	Live blocking investigation	1 query
get_deadlocks	server_name?: str, hours_back?: int	Any time someone says 'we got a deadlock'	1 query
analyze_query_plan	plan_handle: bin, or xml: str	Tuning a specific slow query	1 query + 30 rule evaluations
get_memory_clerks	server_name?: str	Memory pressure investigation	1 query
get_tempdb_usage	server_name?: str	Tempdb filling up or spills in plans	1 query

비용 및 제한

운영 비용

API 쿼터: No external quota; bounded by your monitoring DB capacity
호출당 토큰: Most tools return 200-2000 tokens; plan XML can be 10k+
금액: Free, open-source (Lite and Full editions)
팁: Lite edition is enough for most single-instance shops. Full adds cross-server monitoring.

보안

권한, 시크릿, 파급범위

최소 스코프: VIEW SERVER STATE on the target instance

자격 증명 저장: Monitoring DB connection string managed in the app's settings — stays on your box

데이터 외부 송신: MCP binds to localhost only; no outbound calls from the MCP process

절대 부여 금지: sysadmin or db_owner — VIEW SERVER STATE is enough

Plan XML and query text may contain sensitive data — be careful pasting into shared chat logs.
Run the collector under a dedicated login, not your personal account.

문제 해결

자주 발생하는 오류와 해결

Cannot connect to http://localhost:5151/

The app isn't running or MCP is disabled in Settings. Open the app, enable MCP, and confirm the port.

확인: curl http://localhost:5151/

Login failed for user

Monitoring DB credentials are wrong. Re-enter in the app and confirm VIEW SERVER STATE.

확인: sqlcmd -S <server> -U <user> -P <pwd> -Q 'SELECT 1'

Plan analyzer returns no findings on clearly slow query

Plan may not be in cache. Re-run the query with OPTION(RECOMPILE) and try again.

확인: SELECT plan_handle FROM sys.dm_exec_query_stats WHERE ...

Port 5151 already in use

Change the MCP port in Settings (range 1024-65535) and re-register with your client.

확인: netstat -an | grep 5151

대안

PerformanceMonitor 다른 것과 비교

대안	언제 쓰나	단점/장점
sp_BlitzFirst / Brent Ozar scripts	You prefer running T-SQL procs directly without an app	No MCP — you paste output into the chat yourself
SentryOne / Redgate SQL Monitor	Enterprise environments that already have a commercial monitor	Paid; most don't expose an MCP surface
postgres MCP	You're on Postgres, not SQL Server	Different DB entirely

PerformanceMonitor