AI编程会话自动记录与文档生成工具

vibe-coding-mcp

Your AI coding sessions, documented automatically. Captures every Claude Code session from its transcript — prompts, files changed, commands, design decisions — then turns them into standups, weekly reports, ADRs and PR descriptions, and publishes them to Notion / Obsidian / GitHub Wiki / Confluence / Slack / Discord.

English · 한국어

Why vibe-coding?

You spent four hours pair-programming with Claude. You made three architectural decisions, wrote eight files, and reverted twice. Tomorrow you will not remember why you picked the second option over the first. Most people solve this with manual journaling that lasts two weeks, or by trawling git log after the fact — and git log knows what changed, not why.

vibe-coding-mcp fixes that at the source. Claude Code already writes a full transcript of every session to ~/.claude/projects/. A tiny hook (vibe-coding-mcp hook) parses that transcript after every turn and stores a structured session locally: what you asked, which files were written or edited (with the code), which commands ran, how long it took, which branch, and the design decisions it can detect in the conversation. From there, 17 MCP tools turn those sessions into a standup, a weekly report, a daily log, a DESIGN / ADR / README / API doc, or a PR description, and publish them anywhere.

No API key needed for capture. Everything stays on your disk until you publish.

30-second install

# 1. Register the MCP server with Claude Code (tools, resources, slash-command prompts)
claude mcp add -s user vibe-coding-mcp -- npx -y vibe-coding-mcp

# 2. Install the capture hooks into ~/.claude/settings.json (backup written first)
npx -y vibe-coding-mcp init

# 3. Restart Claude Code. Then import the sessions you already have and check health
npx -y vibe-coding-mcp capture --all
npx -y vibe-coding-mcp doctor

Then, inside Claude Code:

/standup                     → Yesterday / Today / Blockers from captured sessions
/weekly-report               → Shipped / Decisions / In progress / Next, per project
/daily-vibe-log              → today's sessions as one log, offer to publish
@vibe-coding://search/redis  → every session that touched redis (titles, prompts, or file paths)

Or just ask: "why did we choose JWT over sessions last week?", "document yesterday's session as a DESIGN doc and put it in my Obsidian vault", "write the PR description for this refactor".

Upgrading from v2.14? The old README's hook recipe used claude mcp call, which is not a real Claude Code command, so nothing was ever captured. Run npx vibe-coding-mcp init and remove the old entries. See docs/AUTO_CAPTURE.md.

What's new in v2.15.0

• Auto-capture that works. vibe-coding-mcp hook reads Claude Code's hook JSON on stdin, parses the session transcript, and upserts one session per Claude session (idempotent, exit 0 always, nothing on stdout).
• init / doctor / capture CLI. Install the hooks with a merge + backup, diagnose, backfill history.
• muse_digest. Standup / weekly / monthly report over captured sessions, grouped by day · project · tag, English or Korean.
• muse_capture_transcript. Capture on demand from inside the chat ("save this session").
• Resources vibe-coding://sessions/today, vibe-coding://search/{query} and prompts /standup, /weekly-report.
• Agent Skill. skills/vibe-coding/ tells any skill-aware agent (Claude Code, Cursor, Codex, …) when and how to use the server. npm run skill:pack builds the zip.

Real use cases

1. "Write my standup"

/standup calls muse_digest for yesterday and today plus muse_git status, and returns three bullet groups. Every bullet traces to a captured session title, file path, or decision — nothing invented.

2. "What did I ship this week?" (for your 1:1, Slack, or Notion)

/weekly-report calls muse_digest { period: "week", groupBy: "project" }: totals (sessions, hours, files, decisions), highlights (decisions with rationale), per-project sessions with the files they touched, and most-touched files. Optionally muse_publish_document to Notion / Slack.

3. "Why did we pick option B?"

@vibe-coding://search/<keyword> matches titles, summaries, tags and touched file paths. Open the session; designDecisions holds problem → choice → rationale, and summary holds the actual prompts you typed. This is the context a diff cannot give you.

4. "PR description for a refactor that touched 11 files"

/refactor-context chains session → muse_summarize_design_decisions → muse_analyze_code (AST + Mermaid) → muse_git diff into Why / What changed / Trade-offs / How to review.

5. "Fill my Obsidian vault with ADRs"

muse_summarize_design_decisions over a session, muse_generate_dev_document { documentType: "DESIGN" }, muse_publish_document { platform: "obsidian" } — frontmatter, tags and a session back-reference included.

Tools / Resources / Prompts

The skill

skills/vibe-coding/ is an Agent Skill (SKILL.md + references/ + scripts/setup.sh + assets/). It teaches the agent: check that capture is running before reporting, which tool chain matches which request, output shapes for standup / weekly / PR, and the rule that every reported line must trace to a captured session. Install it alongside the MCP server:

# Claude Code: drop it into your skills directory
cp -r skills/vibe-coding ~/.claude/skills/

# or build a zip for a marketplace (Agensi, etc.)
npm run skill:pack      # → dist-skill/vibe-coding-skill-<version>.zip

How it works

 Claude Code ──Stop / SessionEnd hook──▶ vibe-coding-mcp hook ──▶ ~/.vibe-coding-mcp/sessions/*.json
      │   (stdin: session_id,             │ parse ~/.claude/projects/<slug>/<id>.jsonl
      │    transcript_path, cwd)          │ prompts · files+code · commands · timing · decisions
      │                                   │ upsert by claudeSessionId
      ▼
 Claude (or any MCP client) ──▶ src/cli.ts ──▶ src/stdio.ts ─┐
                              src/index.ts (HTTP, POST /mcp) ─┴─▶ src/core/mcpServerFactory.ts
                                                                    ├─ toolRegistry.ts   (17 muse_* tools)
                                                                    ├─ resources.ts      (5 resources)
                                                                    └─ prompts.ts        (5 prompts)
                                                                          │
                                                                          ▼
                                            src/tools/*.ts · src/core/{captureEngine,transcriptParser}.ts
                                            src/core/sessionStorage.ts · src/platforms/*.ts

Design choices that matter:

1. Capture reads the transcript, not the editor. Claude Code already logs everything; we parse it after the fact. The hook does no network calls, writes nothing to stdout, and exits 0 even on failure, so it can never slow down or break a session.
2. One session per Claude session. The Stop hook fires after every turn; captures are keyed on claudeSessionId and update in place. User-added tags survive updates.
3. Local-first. ~/.vibe-coding-mcp/sessions/<id>.json. No DB, no telemetry. grep, back up, or sync it however you like.
4. One registry, two transports. HTTP and stdio expose the same 17 tools / 5 resources / 5 prompts via createMcpServer().

CLI reference

vibe-coding-mcp                 start the MCP server (stdio)
vibe-coding-mcp init            install Stop + SessionEnd hooks into ~/.claude/settings.json
    --project | --local         target .claude/settings.json | .claude/settings.local.json
    --dry-run | --print         show merged file | print only the hooks block
    --remove                    uninstall
    --command "<cmd>"           custom hook command
vibe-coding-mcp capture         capture the newest transcript for the cwd
    --all                       every transcript for the cwd (backfill)
    --transcript <path> --cwd <dir> --dry-run --json
vibe-coding-mcp doctor          hooks · transcripts · store · last hook run · platforms
vibe-coding-mcp hook            (used by the hooks) stdin JSON → capture

Configuration

Tools that need an unconfigured integration return a structured INTEGRATION_NOT_CONFIGURED error instead of crashing.

Known limitations

• Claude Code only (for now). Capture parses Claude Code's transcript format. Cursor / Codex transcripts are on the roadmap. The MCP tools themselves work from any MCP client.
• Decision extraction is heuristic unless useAI: true with an API key: it looks for "decided / chose / instead of / because" patterns in EN and KO.
• AST analysis covers TS / Python / Go.
• Publishing is one-way. Edits made in Notion / Wiki are not pulled back.
• Single-user store. Team search across everyone's sessions needs a shared store (roadmap).

When NOT to use this

• Automatic commit messages → use aicommits or similar; this generates documents.
• Team-wide search across everyone's sessions → build a Notion DB on top, or wait for the shared store.
• Hand-written rich notes → use Obsidian / Notion directly; this is capture-then-publish.

Roadmap

• Cursor / Codex CLI transcript capture.
• Shared session store (S3 / Git / WebDAV) and team mode.
• prepare-commit-msg suggestions from the current session's decisions.
• Rust / Swift / Kotlin AST + diagrams.

Contributing

git clone https://github.com/MUSE-CODE-SPACE/vibe-coding-mcp
cd vibe-coding-mcp
npm install
npm run typecheck && npm test && npm run build
node dist/cli.js capture --dry-run     # try the capture engine on your own transcript

Adding a tool is one file in src/tools/<name>.ts, a Zod schema in src/core/schemas.ts, and one entry in src/core/toolRegistry.ts. CI runs typecheck + build + tests on Node 22 and 24.

Security

Paths are validated, filenames sanitized, webhook URLs must be HTTPS and allowlisted, outbound HTTP has timeouts and backoff — see src/core/security.ts and SECURITY.md. The capture hook stores only your prompts, the assistant's prose, the code it wrote, and the commands it ran; tool outputs, sub-agent side-chains and system reminders are dropped.

Report vulnerabilities privately via a GitHub Security Advisory: https://github.com/MUSE-CODE-SPACE/vibe-coding-mcp/security/advisories/new.

License

MIT — see LICENSE.

Maintainer

@yoon-k (MUSE-CODE-SPACE). Issues + support: https://github.com/MUSE-CODE-SPACE/vibe-coding-mcp/issues.

한국어 요약

vibe-coding-mcp는 Claude Code 세션을 자동으로 문서화하는 MCP 서버 + 에이전트 스킬입니다. Claude Code가 남기는 트랜스크립트(~/.claude/projects/)를 훅이 매 턴마다 파싱해서 프롬프트 · 수정한 파일과 코드 · 실행한 명령 · 소요 시간 · 브랜치 · 설계 결정을 ~/.vibe-coding-mcp/sessions/에 저장합니다. API 키 없이 동작하고, 발행하기 전까지는 아무것도 외부로 나가지 않습니다.

claude mcp add -s user vibe-coding-mcp -- npx -y vibe-coding-mcp   # MCP 등록
npx -y vibe-coding-mcp init                                        # 훅 설치 (settings.json 백업 후 병합)
npx -y vibe-coding-mcp capture --all                               # 기존 세션 가져오기
npx -y vibe-coding-mcp doctor                                      # 상태 확인

Claude Code 안에서 /standup(어제·오늘·블로커), /weekly-report(주간 보고, language: ko 지원), /daily-vibe-log, @vibe-coding://search/키워드를 바로 쓸 수 있고, "지난주에 왜 JWT를 골랐지?", "어제 세션을 DESIGN 문서로 만들어서 Obsidian에 넣어줘", "이 리팩터링 PR 설명 써줘" 같은 요청에 muse_* 도구 17개가 대응합니다.

v2.14의 훅 레시피(claude mcp call)는 존재하지 않는 명령이라 실제로 캡처가 되지 않았습니다. v2.15에서 init으로 교체하세요.

설치·검증·백필은 docs/AUTO_CAPTURE.md, 변경 이력은 CHANGELOG.md, 스킬 패키지는 skills/vibe-coding/ 참고.