vibe-coding-mcp

Auto-documents your AI coding sessions — wire it as a Claude Code hook and your daily README / design-decision log / refactor-PR description writes itself.

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 in commit a1b2c3d over the first one you tried. Most people solve this with manual journaling that lasts two weeks, or by trawling git log after the fact.

vibe-coding-mcp solves it differently: it sits as an MCP server next to your editor, collects code blocks and design decisions out of the conversation, generates README / DESIGN / API / ARCHITECTURE documents from them, and publishes to Notion / GitHub Wiki / Obsidian / Confluence / Slack / Discord — all from inside the chat with Claude. Wire it as a Claude Code PostToolUse + Stop hook (one block of JSON, see Quickstart) and every session auto-captures into ~/.vibe-coding-mcp/sessions/ without you typing anything. The bundled daily-vibe-log prompt then rolls today's captures into one document.

The Phase 3 refactor (v2.14.0, May 2026) collapsed the two-entry-point drift (index.ts had 7 tools, stdio.ts had 15) into a single transport-agnostic registry, so HTTP and stdio now expose the same 15 tools, 3 resources, and 3 prompts.

What's new in v2.14.0

• One registry, two entry points, zero drift. src/core/toolRegistry.ts is the single source of truth — both index.ts (Streamable HTTP) and stdio.ts (bin) load the same 15 tools, 3 resources, and 3 prompts.

• Migrated to high-level McpServer API (SDK 1.25+). McpServer.registerTool() / .resource() / .prompt() instead of low-level setRequestHandler(CallToolRequestSchema). Adding a tool is now a one-line register() call.

• Streamable HTTP (MCP 2025-03-26), not SSE. POST /mcp at :3000. The legacy /sse route returns HTTP 410 with a pointer.

• 3 new MCP Resources for @-mention. vibe-coding://sessions/list (captured sessions), vibe-coding://sessions/{id} (one session in full), vibe-coding://config (current platform creds).

• 3 new MCP Prompts. daily-vibe-log (roll today's sessions into one doc), document-session (one session → README/DESIGN/API/etc. → publish), refactor-context (session → PR description with decisions + AST analysis + git diff).

• +34 tests (149 → 183 passing) covering the registry, resources, prompts, and the McpServer factory.

5-minute Quickstart

# 1. Install via Claude Code
claude mcp add vibe-coding-mcp -- npx -y vibe-coding-mcp

2. Wire it as a Claude Code hook so every session auto-captures. Add this to ~/.claude/settings.json:

{
  "mcpServers": {
    "vibe-coding-mcp": {
      "command": "npx",
      "args": ["-y", "vibe-coding-mcp"]
    }
  },
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Edit|Write|NotebookEdit",
        "hooks": [{
          "type": "command",
          "command": "claude mcp call vibe-coding-mcp muse_session_history '{\"action\":\"save\",\"title\":\"auto-capture\",\"summary\":\"PostToolUse\",\"tags\":[\"auto-capture\"]}' >/dev/null 2>&1 || true"
        }]
      }
    ],
    "Stop": [
      {
        "hooks": [{
          "type": "command",
          "command": "claude mcp call vibe-coding-mcp muse_create_session_log '{\"title\":\"Claude Code session\",\"summary\":\"Stop hook\",\"options\":{\"logType\":\"session\"}}' >/dev/null 2>&1 || true"
        }]
      }
    ]
  }
}

3. Restart Claude Code. The next time you edit a file in a session, capture starts.

4. After a session, try this in Claude:

오늘 캡처된 세션들 합쳐서 daily vibe log 만들어줘. (use the /daily-vibe-log prompt)

5. Expected: Claude calls muse_session_history(action='list', filterTags=['auto-capture']) to gather today's sessions, then muse_create_session_log to compose, then offers to publish to Notion / Obsidian / GitHub Wiki. See docs/AUTO_CAPTURE.md for hook customization (e.g. piping git diff --stat into the capture payload).

Real use cases

1. "I want today's coding session as a daily log without lifting a finger"

Problem: You'll never sit down at the end of the day and write a journal entry. Nobody does. With this MCP: the PostToolUse + Stop hooks save sessions automatically into ~/.vibe-coding-mcp/sessions/. Next morning, run the /daily-vibe-log prompt and Claude reads vibe-coding://sessions/list for today, compiles a Markdown log (problems, decisions, code blocks, blockers), and offers to publish to Notion. Why it's better than git log or Notion templates: git log knows what you committed, not why you picked option B over option A. The session capture preserves the conversation context that the diff doesn't.

2. "I need a PR description for a refactor that touched 11 files"

Problem: Writing a good PR description for a refactor takes 15 minutes of re-reading the diff. Most refactor PRs end up with a one-liner and reviewers hate it. With this MCP: invoke the /refactor-context prompt against the session you just finished. The prompt chains muse_session_history(get) → muse_summarize_design_decisions → muse_analyze_code (AST + Mermaid diagram) → muse_git(diff) and outputs a structured PR description with: Why (decisions), What changed (file-by-file from AST), Trade-offs (decisions section), and How to review (diagram of new architecture). Why it's better than asking Claude to "write a PR description": the prompt explicitly pulls the design-decision summary out of the session, not just the diff — reviewers get the why, not a paraphrased diff.

3. "I want my Obsidian vault to fill itself with my design decisions"

Problem: You make architectural decisions during pair-programming and they evaporate. ADR templates require manual discipline you don't have. With this MCP: call muse_summarize_design_decisions over your session, then muse_publish_document({ platform: 'obsidian', vault: '~/MyVault/decisions' }). The Obsidian platform writer adds YAML frontmatter (tags, date, session-id back-reference), so the new note shows up correctly in your vault graph view. Why it's better than a Notes/ADR-template.md: the decision text is generated from the actual code+conversation, not from your tired post-session memory. And it works across 6 platforms (Notion, GitHub Wiki, Obsidian, Confluence, Slack, Discord) with one tool call.

Tools / Resources / Prompts

How it works

                      Claude (or any MCP client)
                                │
       ┌────────────────────────┴────────────────────────┐
       ▼                                                  ▼
  src/index.ts (Streamable HTTP, POST /mcp at :3000)   src/stdio.ts (bin)
       │                                                  │
       └────────────────────────┬─────────────────────────┘
                                ▼
                  src/core/mcpServerFactory.ts
                                │
              ┌─────────────────┼─────────────────┐
              ▼                 ▼                 ▼
         toolRegistry.ts   resources.ts      prompts.ts
         (15 muse_* tools) (3 resources)     (3 prompts)
              │
              ▼
              src/tools/*.ts (15 tool implementations)
                            │
                            ▼
              src/core/sessionStorage.ts (~/.vibe-coding-mcp/sessions/)
              src/platforms/*.ts (notion / github-wiki / obsidian / ...)

Three design choices that matter:

1. Single registry, two transports. HTTP and stdio both call createMcpServer() from mcpServerFactory.ts. There's no "stdio has feature X that HTTP doesn't" — the v2.13.0 drift is closed.

2. Sessions are local files. ~/.vibe-coding-mcp/sessions/<id>.json — no remote DB, no telemetry, you own the data. Trivial to back up, grep, or sync via Dropbox.

3. Publishing is per-platform, not one-size-fits-all. Notion expects blocks, GitHub Wiki expects sidebar markdown, Obsidian expects frontmatter — muse_normalize_for_platform handles each one's quirks so muse_publish_document can stay a single tool.

Configuration

All env vars are optional. The MCP boots with zero env vars set — tools that require an integration return a structured INTEGRATION_NOT_CONFIGURED error instead of crashing.

Known limitations

• Capture is hook-driven. The MCP itself doesn't sniff your editor — it relies on Claude Code calling it (via the PostToolUse / Stop hooks shown in Quickstart). Without hooks, you have to call muse_session_history manually.

• No shared session store yet. Sessions are local. Two laptops = two session histories. Cross-device sync is on the roadmap.

• AST analysis covers TS / Python / Go. Rust, Swift, Kotlin, Ruby AST support is not in v2.x.

• Publishing is one-way. We can write to Notion / Wiki / Obsidian etc. but we don't pull updates back into the session store.

When NOT to use this

• If you want automatic git commit messages, use aicommits or similar — vibe-coding-mcp generates documents, not commit messages.

• If you want team-wide knowledge base search across everyone's sessions, this MCP is single-user/local. Build a Notion DB on top and search there, or wait for the shared-store roadmap item.

• If you want a manual journaling tool with rich editing, use Obsidian / Notion directly — this MCP is for the auto-capture-then-publish flow, not for hand-writing notes.

Comparison

Roadmap

• Shared session store. Opt-in cross-device sync (S3 / Git / WebDAV) and team mode where multiple users can search each other's captured sessions.

• Auto-summary by topic. Group sessions by topic across weeks ("everything I did on the auth refactor in Q2") instead of by day.

• Git commit-message integration. Surface muse_summarize_design_decisions output as a prepare-commit-msg suggestion.

• Wider AST language support. Rust + Swift + Kotlin AST + Mermaid diagrams.

• Resource for cross-session search. vibe-coding://search?q=... so the LLM can @-mention "all sessions about caching" in one go.

Contributing

PRs welcome. Adding a tool is now a single-file change in src/tools/<name>.ts + a one-line register() in src/core/toolRegistry.ts. CI runs typecheck + build + test on Node 20 against every PR.

Quick contributor loop:

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

Security

This package validates paths, sanitizes filenames, enforces HTTPS-only + allowlisted hosts for webhook URLs, and wraps outbound HTTP with a timeout + exponential-backoff retry — see src/core/security.ts. The threat model, supported versions, and vulnerability reporting process are in SECURITY.md. CodeQL with the security-and-quality query pack runs on every push and weekly.

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

License

MIT — SPDX identifier declared in package.json (a top-level LICENSE file will be added in the next release).

Maintainer

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

한국어 요약

vibe-coding-mcp는 Claude(또는 다른 MCP 클라이언트)와 함께 AI 페어 코딩 세션을 자동으로 문서화해 주는 MCP 서버입니다. Claude Code의 PostToolUse + Stop 훅에 연결하면 매 세션이 ~/.vibe-coding-mcp/sessions/ 에 자동 저장되고, 15개 도구가 거기서 README / DESIGN / API / ARCHITECTURE / CHANGELOG / TUTORIAL 6종 문서를 생성한 뒤 Notion / GitHub Wiki / Obsidian / Confluence / Slack / Discord 6개 플랫폼에 발행합니다.

v2.14.0(2026-05-20)에서 HTTP/stdio 두 진입점이 단일 레지스트리 를 공유하도록 통합되었고, @-mention 가능한 MCP Resources 3개(sessions/list, sessions/{id}, config)와 Prompts 3개(daily-vibe-log, document-session, refactor-context)가 추가되었습니다.

설치 + 자동 캡처 훅 설정은 docs/AUTO_CAPTURE.md 참고. 변경 이력은 CHANGELOG.md, 위협 모델은 SECURITY.md.