MCP Daemon Isolation for Query-Type MCPs

External CLI pattern for isolating query-type MCP tool results from main context.

Query-Type MCP 정의

| Type | Examples | 특징 | |------|----------|------| | Query-type | Serena (LSP), Database, Search | 결과 크기 예측 불가, 수천 토큰 가능 | | Action-type | File write, Git, Deploy | 결과 작음, 성공/실패 위주 |

식별 기준: find_*, search_*, get_*, list_* 패턴 도구

Problem

Claude Session
├── mcp__serena__find_symbol("UserService")
│   └── Result: 2,500 tokens (full JSON)    ← CONSUMED
└── Context budget: rapidly depleting

Daemon isolates tool definitions (~350 tokens × tools), but results still consume context.

Solution: External CLI + Structured Extraction

Claude Session
├── Bash: serena-query find_symbol UserService
│   └── stdout: "• UserService [Class] @ src/user.py:42"
│       (~50 tokens)                      ← MINIMAL
└── Full result stored: /tmp/serena-result.json

Key insight: Claude IS the LLM. Structured extraction provides 95%+ token savings.

Architecture

┌──────────────┐     ┌──────────────────┐     ┌──────────────┐
│ Claude       │────▶│ serena-query     │────▶│ Serena       │
│ Session      │     │ (external CLI)   │     │ Daemon :8765 │
└──────────────┘     └──────────────────┘     └──────────────┘
     Bash (~50 tok)     SSE + MCP Protocol       29 tools

For protocol details: Read("references/mcp-sse-protocol.md")

Structured Extractors

| Tool | Full JSON | Extracted Output | Savings | |------|-----------|------------------|---------| | list_dir | ~800 | 📁 Dirs(12): hooks... 📄 Files(11) | 95% | | get_symbols_overview | ~1,200 | Class(2): Config... Function(5) | 96% | | find_symbol | ~2,000 | • UserService [Class] @ src/user.py:42-98 | 97% | | search_for_pattern | ~1,500 | Matches(15) in 8 files | 95% |

For formatter code: Read("references/extractors-and-examples.md")

Usage

Output Modes

serena-query find_symbol UserService                    # summary (기본)
serena-query find_symbol UserService --mode location    # 위치만 (Read 연계)
serena-query find_symbol UserService --mode full        # 전체 JSON

Basic Commands

serena-query list_dir .
serena-query get_symbols_overview src/main.py --depth 1
serena-query find_symbol UserService --path src/
serena-query search_for_pattern "class.*Service" --path src/
serena-query find_symbol UserService --output /tmp/result.json

최적 워크플로우: 탐색-위치-읽기

1. 탐색 (--mode summary)     ~25 토큰   "어떤 심볼이 있지?"
2. 위치 (--mode location)    ~15 토큰   "정확히 어디있지?"
3. 읽기 (Read 도구)          실제 크기   필요한 코드만

| 시나리오 | stdio 방식 | daemon+Read | 절감률 | |---------|-----------|-------------|--------| | 클래스 분석 | ~525 | ~240 | 54% | | 11개 함수 분석 | ~4,300 | ~665 | 85% | | 대규모 검색 | ~10,000+ | ~500 | 95% |

For detailed examples: Read("references/extractors-and-examples.md")

Installation

pip install httpx
cp scripts/serena-query ~/.local/bin/
chmod +x ~/.local/bin/serena-query

Daemon Setup

uvx --from git+https://github.com/oraios/serena \
  serena start-mcp-server --transport sse --port 8765 --project-from-cwd

Decision: Structured vs LLM Summarization

| Aspect | LLM Summarization | Structured Extraction | |--------|-------------------|----------------------| | Latency | +2-5 seconds | ~0ms | | Cost | API call per query | Zero | | Consistency | Variable | Deterministic |

Claude IS the LLM consuming the output - no need for additional summarization.

격리 전략 가이드

| 시나리오 | 권장 | 이유 | |---------|------|------| | 단일 심볼 편집 | stdio | 즉각적인 수정 | | 코드베이스 탐색 | daemon | 대량 결과 예상 | | 참조 추적 | daemon + location | 위치만 필요 | | 리팩토링 계획 | daemon + full | 전체 구조 분석 |

Common Issues

| 증상 | 원인 | 해결 | |------|------|------| | "Received request before initialization" | 핸드셰이크 누락 | initialize → notifications/initialized → tools/call 순서 | | "Connection refused on 8765" | Daemon 미실행 | systemctl --user start serena-daemon | | "Empty result" | 구조 불일치 | --output으로 raw JSON 저장 후 확인 |

References

Implementation

Protocol & Examples

MCP Gateway Patterns

mcp-daemon-isolation