concierge MCP — 사용 사례, 설치 & 라이브 데모

왜 쓰나요

핵심 기능

Progressive disclosure — only expose tools relevant to the current workflow step
Workflow state shared across steps without manual plumbing
Semantic tool search — worth it once you cross ~100 tools
Multiple transports: stdio, HTTP, SSE
Session-isolated — concurrent workflows don't clobber each other

라이브 데모

실제 사용 모습

concierge.replay ▶ 준비됨

0/0

설치

클라이언트 선택

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

{
  "mcpServers": {
    "concierge": {
      "command": "uvx",
      "args": [
        "concierge"
      ],
      "_inferred": true
    }
  }
}

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

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "concierge": {
      "command": "uvx",
      "args": [
        "concierge"
      ],
      "_inferred": true
    }
  }
}

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

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "concierge": {
      "command": "uvx",
      "args": [
        "concierge"
      ],
      "_inferred": true
    }
  }
}

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

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "concierge": {
      "command": "uvx",
      "args": [
        "concierge"
      ],
      "_inferred": true
    }
  }
}

Claude Desktop과 같은 형식. Windsurf 재시작 후 적용.

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "concierge",
      "command": "uvx",
      "args": [
        "concierge"
      ]
    }
  ]
}

Continue는 맵이 아닌 서버 오브젝트 배열 사용.

~/.config/zed/settings.json

{
  "context_servers": {
    "concierge": {
      "command": {
        "path": "uvx",
        "args": [
          "concierge"
        ]
      }
    }
  }
}

context_servers에 추가. 저장 시 Zed가 핫 리로드.

claude mcp add concierge -- uvx concierge

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

사용 사례

실전 활용법: concierge

Wrap a giant REST API (200+ endpoints) as an MCP without blowing the context

👤 Platform engineers exposing internal APIs to LLMs ⏱ ~90 min advanced

언제 쓸까: Your company's API has 300 endpoints and a naive MCP dumps all of them into the system prompt, wrecking latency and hit rate.

사전 조건

Python 3.9+ — pyenv / uv
OpenAPI spec or endpoint catalog — Your API gateway / Swagger

흐름

Scaffold with concierge-sdk

Using concierge-sdk, scaffold an MCP server that wraps my OpenAPI spec at ./openapi.yaml. Make it use semantic search rather than listing all tools up front.✓ 복사됨

→ Boilerplate code + search handler
Define workflow stages

Group the endpoints into 3 workflows: 'read operations', 'create operations', 'admin'. Each workflow exposes only its own tools.✓ 복사됨

→ Workflow definitions with tool allowlists
Test with Claude

Connect Claude Desktop to this server and verify that listing tools only shows the search tool + current workflow tools — not all 300.✓ 복사됨

→ Claude sees ~10 tools, not 300

결과: A large API surface usable by an LLM without collapsing on system prompt length.

함정

Semantic search returns the wrong tool when descriptions are too similar — Write distinctive one-line tool descriptions; test search with held-out queries

Build a guided multi-step workflow (e.g. customer onboarding)

👤 Devs building structured agent flows ⏱ ~60 min advanced

언제 쓸까: You want the LLM to follow a defined sequence: collect info → validate → create record → notify. Each step has its own tools.

흐름

Declare the workflow

Define a concierge workflow 'customer_onboarding' with steps [collect, validate, create, notify], each with its own tool set.✓ 복사됨

→ Workflow config
Share state

Pass the customer_data dict from step 1 into steps 2, 3, 4 via shared state. Show me how.✓ 복사됨

→ State-object code
Handle failure

If validate fails, return to step 1 with the specific fix needed.✓ 복사됨

→ Retry/backtrack logic

결과: A robust guided flow that keeps the LLM on-rails.

함정

LLM tries to skip steps — Concierge enforces order at the tool-visibility level — skipping is physically impossible if you wired it right

Ship your first MCP server in 15 minutes

👤 Developers new to MCP ⏱ ~15 min beginner

언제 쓸까: You want to expose 3-5 functions to Claude and don't want to learn the raw MCP spec.

흐름

Install and scaffold

pip install concierge-sdk. Generate a minimal server exposing two tools: add(a, b) and greet(name). Stdio transport.✓ 복사됨

→ Working Python file
Run and connect

Add to Claude Desktop config and test that both tools are callable.✓ 복사됨

→ Tool calls succeed in Claude

결과: A working MCP server you wrote end-to-end in an afternoon.

함정

Missing type hints — SDK relies on them for tool schema — Always type your args and return — concierge generates the MCP schema from them

조합

다른 MCP와 조합해 10배 효율

concierge + gateway

Serve concierge-built MCP behind a security gateway for PII redaction

Start concierge server on :8001, then add it as an upstream to mcp-gateway with Presidio plugin.✓ 복사됨

도구

이 MCP가 노출하는 것

도구	입력	언제 호출	비용
concierge.workflow(name)	name, steps, initial_state	Define a named multi-step flow	framework only
concierge.tool(workflow, step)	step allowlist	Attach a function as a tool scoped to one or more steps	framework only
concierge.search_tools	query: str	Auto-exposed when you have too many tools to list eagerly	free
concierge.serve()	transport: 'stdio'\|'http'\|'sse'	Entry point of your script	free

비용 및 제한

운영 비용

API 쿼터: None — you're the author
호출당 토큰: Depends on your tool design; progressive disclosure shrinks system prompts 5-10x
금액: Free and open source
팁: Don't eagerly expose every tool. Start with 'search' + current-step tools. Add eager tools only if profiling shows the LLM always needs them.

보안

권한, 시크릿, 파급범위

자격 증명 저장: Your MCP server handles its own credentials — concierge doesn't impose a model

데이터 외부 송신: Whatever your tools call

As an SDK, concierge's threat model is inherited from how you use it. Validate inputs in each tool.
Session isolation protects state per-user but shared external resources still need their own locks.

문제 해결

자주 발생하는 오류와 해결

Tools not appearing in the expected step

Check the step arg on @concierge.tool. A tool with step=['create'] is invisible during collect.

확인: Call the search_tools tool; if nothing returns, your allowlist is wrong

Semantic search returns no results

Tool descriptions may be empty. Concierge builds embeddings from docstrings — fill them in.

Workflow state lost between calls

Concierge state is session-scoped. Ensure your client is sticky (same MCP session across calls).

대안

concierge 다른 것과 비교

대안	언제 쓰나	단점/장점
fastmcp (Python)	You want the most popular Python MCP SDK without progressive disclosure	Lighter, but no built-in workflow/tool-gating features
Official MCP Python SDK	You want zero framework, closest to the spec	More boilerplate; you build gating yourself
TypeScript MCP SDK	Your stack is TS/Node	Different language; no direct concierge equivalent

concierge