inspector MCP — 사용 사례, 설치 & 라이브 데모

Name: inspector MCP Server
Author: MCPJam

왜 쓰나요

핵심 기능

Full JSON-RPC tracing across host/server
Inline LLM chat while inspecting tool calls and context
Evaluation runs across multiple models
OAuth flow debugging with conformance checks
Workspace sharing, CLI, SDK, and CI/CD integration

라이브 데모

실제 사용 모습

inspector.replay ▶ 준비됨

0/0

설치

클라이언트 선택

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

{
  "mcpServers": {
    "inspector": {
      "command": "npx",
      "args": [
        "-y",
        "inspector"
      ],
      "_inferred": true
    }
  }
}

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

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "inspector": {
      "command": "npx",
      "args": [
        "-y",
        "inspector"
      ],
      "_inferred": true
    }
  }
}

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

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "inspector": {
      "command": "npx",
      "args": [
        "-y",
        "inspector"
      ],
      "_inferred": true
    }
  }
}

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

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "inspector": {
      "command": "npx",
      "args": [
        "-y",
        "inspector"
      ],
      "_inferred": true
    }
  }
}

Claude Desktop과 같은 형식. Windsurf 재시작 후 적용.

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "inspector",
      "command": "npx",
      "args": [
        "-y",
        "inspector"
      ]
    }
  ]
}

Continue는 맵이 아닌 서버 오브젝트 배열 사용.

~/.config/zed/settings.json

{
  "context_servers": {
    "inspector": {
      "command": {
        "path": "npx",
        "args": [
          "-y",
          "inspector"
        ]
      }
    }
  }
}

context_servers에 추가. 저장 시 Zed가 핫 리로드.

claude mcp add inspector -- npx -y inspector

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

사용 사례

실전 활용법: inspector

How to debug why your MCP tool isn't being called

👤 MCP server developers ⏱ ~15 min intermediate

언제 쓸까: You wrote a tool, Claude doesn't use it, and you don't know why.

사전 조건

Node — nodejs.org or nvm

흐름

Launch Inspector

npx @mcpjam/inspector@latest and open the printed URL.✓ 복사됨

→ UI opens in browser
Connect your server

In the UI, add your server (stdio or HTTP) and connect.✓ 복사됨

→ Tools list populates
Simulate a prompt

In the Chat tab, ask a question that should trigger your tool. Watch the JSON-RPC trace.✓ 복사됨

→ Either tool fires (good) or you see why the model rejected it (bad description, missing input schema)

결과: Concrete reason (wrong tool description, schema issue, model choice) with a visible fix.

함정

Testing with a weaker model that never uses tools — Test across multiple models in the Evals tab — Haiku, Sonnet, GPT-4o — behavior varies

How to run eval tests on PRs for your MCP server

👤 Teams maintaining production MCPs ⏱ ~30 min advanced

언제 쓸까: You want to catch 'oops that prompt no longer picks my tool' in CI.

흐름

Author evals

In Inspector, create eval cases: (prompt, expected_tool_called, expected_args).✓ 복사됨

→ Eval suite saved
Wire into CI

Use the CLI to run evals on every PR; fail the build on regression.✓ 복사됨

→ .github/workflows/mcp-eval.yml runs evals

결과: PRs that break tool selection don't merge.

함정

Evals are flaky when temperature > 0 — Use temperature 0 in eval runs

함께 쓰기: github

How to diagnose a broken OAuth flow on your MCP server

👤 MCP server devs shipping OAuth-protected tools ⏱ ~20 min advanced

언제 쓸까: Tokens aren't refreshing, or redirect returns 400.

흐름

Enable OAuth debug in Inspector

Configure your server's OAuth endpoints and click 'Test flow'.✓ 복사됨

→ Step-by-step trace of authorize, token, refresh
Identify the break

Inspector highlights nonconforming steps (wrong content-type, missing PKCE, bad scope).✓ 복사됨

→ Specific failure node

결과: Conforming OAuth implementation.

조합

다른 MCP와 조합해 10배 효율

inspector + github

Run Inspector evals in a GitHub Action on PRs for your MCP server

Set up a GitHub Action using @mcpjam/inspector CLI to run evals on every PR touching src/tools/.✓ 복사됨

도구

이 MCP가 노출하는 것

도구	입력	언제 호출	비용
(inspector app) connect-server	transport: stdio\|http, command or url	First step	free
(inspector app) execute-tool	tool, args	Manual testing	depends on tool
(inspector app) chat-and-inspect	model, prompt	End-to-end simulation	1 LLM call
(inspector app) run-evals	suite_id	CI regression testing	N LLM calls

비용 및 제한

운영 비용

API 쿼터: Your LLM/eval keys determine the budget
호출당 토큰: Evals can burn tokens — limit concurrency + size
금액: Inspector itself is free. Hosted app at app.mcpjam.com has tiers.
팁: Run evals on Haiku for cheap smoke tests; reserve Sonnet/Opus for nightly runs.

보안

권한, 시크릿, 파급범위

최소 스코프: Whatever scopes your MCP server needs LLM provider keys for evals

자격 증명 저장: Local env for CLI; workspace secrets for the hosted app

데이터 외부 송신: Tool calls stay between your MCP and LLM; hosted app stores eval results in your workspace

절대 부여 금지: Don't upload production OAuth client secrets to the hosted app without review

Inspector can call write tools during eval runs — use a sandbox server, not production.

문제 해결

자주 발생하는 오류와 해결

Can't connect to my stdio server

Make sure the command Inspector runs matches what Claude Desktop runs — working dir and env vars included.

Chat tab says 'model not configured'

Add your Anthropic/OpenAI key in Settings.

Evals pass locally, fail in CI

Env mismatch — pin the same model + temperature in CI as locally.

OAuth test reports 'state mismatch'

Your server isn't preserving the state param; compare to the spec in Inspector's conformance report.

대안

inspector 다른 것과 비교

대안	언제 쓰나	단점/장점
@modelcontextprotocol/inspector (official)	You want the first-party reference inspector	Smaller feature surface; no evals/CI
mcp-client-for-ollama	You want a terminal-only Ollama-backed client	Not a debugger, just a client

inspector