drift MCP — 사용 사례, 설치 & 라이브 데모

왜 쓰나요

핵심 기능

Pattern detection: auto-infers conventions from your actual code
Decision memory: write-once notes that persist across sessions
Multi-language: TypeScript, Rust, Java, Python, C#, PHP
MCP + offline CLI — same engine, two interfaces
Local-first: all data stored on your disk, nothing uploaded

라이브 데모

실제 사용 모습

drift.replay ▶ 준비됨

0/0

설치

클라이언트 선택

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

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

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

~/.cursor/mcp.json · .cursor/mcp.json

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

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

VS Code → Cline → MCP Servers → Edit

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

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

~/.codeium/windsurf/mcp_config.json

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

Claude Desktop과 같은 형식. Windsurf 재시작 후 적용.

~/.continue/config.json

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

Continue는 맵이 아닌 서버 오브젝트 배열 사용.

~/.config/zed/settings.json

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

context_servers에 추가. 저장 시 Zed가 핫 리로드.

claude mcp add drift -- uvx drift

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

사용 사례

실전 활용법: drift

Onboard an AI assistant to your codebase's conventions

👤 Tech leads tired of reminding Claude 'no, we don't use classes here' ⏱ ~20 min beginner

언제 쓸까: First session on a new repo, or after a big refactor that changed conventions.

사전 조건

drift installed — uvx drift or install globally

흐름

Run the initial scan

Run drift scan on this repo. Tell me what patterns you detected around error handling, module structure, and naming.✓ 복사됨

→ Concrete convention summary with examples
Record decisions you want preserved

Record these team decisions: we use Result<T,E> not exceptions, one controller per feature folder, snake_case for file names. Tag them 'conventions'.✓ 복사됨

→ Confirmation; entries queryable later
Verify future sessions pick it up

In a new session: what conventions does this repo follow?✓ 복사됨

→ Claude recites the decisions you recorded

결과: Every future AI session starts with your team's conventions loaded — fewer corrections, more productive chats.

함정

Patterns are inferred from existing code, including bad code — Review the initial scan and delete any 'conventions' that are actually tech debt you're trying to escape
Decision memory drifts from reality if you forget to update it — Treat memory entries like docs — review quarterly

함께 쓰기: filesystem · github

Enforce team conventions during PR review

👤 Code reviewers ⏱ ~10 min intermediate

언제 쓸까: Before approving a PR, check it didn't silently break team patterns.

흐름

Load this PR's diff

Load the diff for PR #213 and compare it against the conventions drift has recorded.✓ 복사됨

→ List of convention-matches and misses
Draft review comments for deviations

For each deviation, draft a polite review comment citing the convention.✓ 복사됨

→ Comment text per finding

결과: Consistent PR reviews without the reviewer having to remember every unwritten rule.

함정

New patterns may be intentional — don't be dogmatic — Let the author override with a 'new pattern' decision that gets recorded for future PRs

함께 쓰기: github

Maintain a lightweight architecture decision log

👤 Any team that keeps forgetting why choices were made ⏱ ~5 min beginner

언제 쓸까: Instead of a formal ADR folder no one updates.

흐름

Record a decision when made

Record decision: we picked Postgres over DynamoDB because of ad-hoc query needs. Date today. Tags: db, architecture.✓ 복사됨

→ Entry saved with ID
Query later when the question returns

Why did we pick Postgres?✓ 복사됨

→ The stored decision surfaces

결과: Institutional memory that survives turnover.

조합

다른 MCP와 조합해 10배 효율

drift + filesystem

drift knows the conventions, filesystem does the edits

Using the conventions drift has recorded, refactor src/api/users.ts to match. Use filesystem to apply edits.✓ 복사됨

drift + github

Review incoming PRs against recorded conventions

Fetch PR #88, check it against drift conventions, draft review comments for any drift.✓ 복사됨

도구

이 MCP가 노출하는 것

도구	입력	언제 호출	비용
scan_codebase	path: str	First session, or after large refactors	free; CPU-bound
list_conventions	tag?: str	Load context at session start	free
record_decision	title: str, body: str, tags?: str[]	When team agrees on something non-obvious	free
query_memory	query: str	When Claude should recall past context	free (local embeddings)
detect_pattern	area: str	Ad-hoc 'how do we normally do X here?'	free

비용 및 제한

운영 비용

API 쿼터: None — fully local
호출당 토큰: Convention summaries: ~1k tokens. Full scan dumps can be 10k+
금액: Free and open source
팁: Load summary conventions at session start, not the full scan. Full scan only on refactor.

보안

권한, 시크릿, 파급범위

자격 증명 저장: Local SQLite or filesystem only

데이터 외부 송신: None by default — runs fully offline

Decision memory may contain sensitive business context. Add the drift data dir to .gitignore.
If you enable any cloud sync features, review what's leaving the machine.

문제 해결

자주 발생하는 오류와 해결

Scan hangs on a large monorepo

Exclude generated and vendor directories via .driftignore (same syntax as .gitignore).

확인: drift scan --dry-run

Memory query returns nothing relevant

Rebuild local embeddings index; new decisions don't surface until indexed.

확인: drift reindex

Claude doesn't use recorded conventions

Make sure the MCP is listed first in your client config so its context loads at the start of every session.

확인: claude mcp list

대안

drift 다른 것과 비교

대안	언제 쓰나	단점/장점
MARM-Systems	You want general cross-session memory not specific to code	Less code-aware; more general notes
llm-context.py	You want rule-driven code bundling rather than memory	No persistence layer
Plain ADRs in the repo	You prefer docs-as-code	AI has to read them each session — higher token cost

drift