swarmclaw MCP — 사용 사례, 설치 & 라이브 데모

왜 쓰나요

핵심 기능

23+ LLM provider integrations under one roof
Delegates to Claude Code, Codex, OpenCode, native subagents
Memory: hybrid recall + graph traversal + durable docs
Schedules, background jobs, heartbeat loops for long-running autonomy

라이브 데모

실제 사용 모습

swarmclaw.replay ▶ 준비됨

0/0

설치

클라이언트 선택

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

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

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

~/.cursor/mcp.json · .cursor/mcp.json

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

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

VS Code → Cline → MCP Servers → Edit

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

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

~/.codeium/windsurf/mcp_config.json

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

Claude Desktop과 같은 형식. Windsurf 재시작 후 적용.

~/.continue/config.json

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

Continue는 맵이 아닌 서버 오브젝트 배열 사용.

~/.config/zed/settings.json

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

context_servers에 추가. 저장 시 Zed가 핫 리로드.

claude mcp add swarmclaw -- npx -y swarmclaw

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

사용 사례

실전 활용법: swarmclaw

Run an always-on research agent with SwarmClaw

👤 Researchers, competitive-intel analysts ⏱ ~45 min advanced

언제 쓸까: You want an agent that wakes up daily, runs a research loop, and drops results somewhere.

사전 조건

SwarmClaw installed — One-click desktop installer or npm i -g @swarmclawai/swarmclaw
First-run access key — Generated on initial launch at http://localhost:3456

흐름

Define the agent

Create a SwarmClaw agent 'daily-ai-news' with provider=Claude, memory=enabled, schedule='every day 08:00'.✓ 복사됨

→ Agent registered
Attach tools via MCP

Attach firecrawl + notion MCPs so the agent can scrape and write results to Notion.✓ 복사됨

→ Tools wired
Let it run

Enable the schedule. Review memory weekly to see what it's learned.✓ 복사됨

→ Daily Notion pages appear

결과: A background research teammate.

함정

LLM cost spiral on long loops — Set per-session token budgets in the agent config; alert on overrun

함께 쓰기: firecrawl

Delegate coding tasks across multiple models with SwarmClaw

👤 Engineers who want quality-vs-speed tradeoffs per sub-task ⏱ ~60 min advanced

언제 쓸까: A feature has planning (Opus), bulk coding (Sonnet), and polish (Codex) phases.

흐름

Define the pipeline

Create a workflow: planner=Claude Opus, coder=Claude Sonnet via Claude Code, reviewer=Codex. Handoffs via transcript.✓ 복사됨

→ Workflow visible in UI
Run on a task

Run the workflow on feature spec [paste]. Watch the handoffs.✓ 복사됨

→ Final PR proposal after each stage contributes

결과: Right tool for each phase of a single feature.

함께 쓰기: github

Build a memory-rich personal assistant with SwarmClaw

👤 Power users who dislike re-explaining context each session ⏱ ~30 min intermediate

언제 쓸까: Your ChatGPT/Claude sessions lose context you wanted retained across days.

흐름

Enable hybrid memory

Create a long-lived session 'assistant' with hybrid memory (vector + graph). Pipe all prior transcripts into the memory.✓ 복사됨

→ Memory populated
Ask and let it recall

What were my open questions from last Tuesday's session about pricing?✓ 복사됨

→ Recalled items with citations to source sessions

결과: An assistant that actually remembers.

조합

다른 MCP와 조합해 10배 효율

swarmclaw + claude-code

SwarmClaw orchestrates; Claude Code is the coder in the loop

Delegate the 'implement-feature' step to Claude Code via the swarmclaw workflow.✓ 복사됨

swarmclaw + firecrawl

Scheduled scraping + memory-enriched analysis

Daily agent: scrape competitor pages, diff against memory, flag changes.✓ 복사됨

도구

이 MCP가 노출하는 것

도구	입력	언제 호출	비용
create_agent	name, provider, memory?, schedule?	Bootstrap a new agent	0
attach_mcp	agent_id, mcp_config	Give the agent tools	0
run_session	agent_id, input	Ad-hoc invocation	LLM tokens
schedule_agent	agent_id, cron	Long-running autonomy	0
memory_query	agent_id, query	Inspect agent memory	0

비용 및 제한

운영 비용

API 쿼터: Whatever your LLM providers enforce
호출당 토큰: Multi-turn sessions with memory retrieval: 3-30k per step
금액: Free self-hosted; LLM token costs are yours
팁: Set hard per-session budgets; autonomous loops can spend unnoticed

보안

권한, 시크릿, 파급범위

자격 증명 저장: LLM keys encrypted in the local data dir; first-run access key protects the UI

데이터 외부 송신: To each configured LLM provider + attached MCPs

절대 부여 금지: internet-exposing the UI without auth — default is localhost only

문제 해결

자주 발생하는 오류와 해결

UI inaccessible after install

SwarmClaw binds to 127.0.0.1:3456; ensure nothing else owns that port. lsof -i :3456

Agent loops spend tokens in the background

Disable its schedule and inspect last transcripts; set a token budget on the agent before re-enabling

MCP server fails to start under swarmclaw

Test the MCP command manually first (with f/mcptools); swarmclaw uses the same invocation

확인: mcp tools -- <your-mcp-cmd>

대안

swarmclaw 다른 것과 비교

대안	언제 쓰나	단점/장점
n8n-workflow-builder	You want deterministic workflow automation not agent loops	Less agentic; no LLM memory
LangGraph / Autogen	You want code-first orchestration in your own app	You host + write everything

swarmclaw