/ Каталог / Песочница / vibe-check-mcp-server
● Сообщество PV-Bhat ⚡ Сразу

vibe-check-mcp-server

автор PV-Bhat · PV-Bhat/vibe-check-mcp-server

A second LLM that challenges your agent's plan before execution — catches tunnel vision, over-engineering, reasoning lock-in.

vibe-check-mcp (PV-Bhat) provides metacognitive feedback to agents via Chain-Pattern Interrupts (CPI). Invokes a second model to push back on assumptions at key decision points. Project is in maintenance mode (v2.8.0 final) but still fully functional under MIT.

Зачем использовать

Ключевые функции

Живое демо

Как выглядит на практике

vibe-check.replay ▶ готово
0/0

Установка

Выберите клиент

~/Library/Application Support/Claude/claude_desktop_config.json  · Windows: %APPDATA%\Claude\claude_desktop_config.json
{
  "mcpServers": {
    "vibe-check": {
      "command": "npx",
      "args": [
        "-y",
        "vibe-check-mcp-server"
      ],
      "_inferred": true
    }
  }
}

Откройте Claude Desktop → Settings → Developer → Edit Config. Перезапустите после сохранения.

~/.cursor/mcp.json · .cursor/mcp.json
{
  "mcpServers": {
    "vibe-check": {
      "command": "npx",
      "args": [
        "-y",
        "vibe-check-mcp-server"
      ],
      "_inferred": true
    }
  }
}

Cursor использует ту же схему mcpServers, что и Claude Desktop. Конфиг проекта приоритетнее глобального.

VS Code → Cline → MCP Servers → Edit
{
  "mcpServers": {
    "vibe-check": {
      "command": "npx",
      "args": [
        "-y",
        "vibe-check-mcp-server"
      ],
      "_inferred": true
    }
  }
}

Щёлкните значок MCP Servers на боковой панели Cline, затем "Edit Configuration".

~/.codeium/windsurf/mcp_config.json
{
  "mcpServers": {
    "vibe-check": {
      "command": "npx",
      "args": [
        "-y",
        "vibe-check-mcp-server"
      ],
      "_inferred": true
    }
  }
}

Тот же формат, что и Claude Desktop. Перезапустите Windsurf для применения.

~/.continue/config.json
{
  "mcpServers": [
    {
      "name": "vibe-check",
      "command": "npx",
      "args": [
        "-y",
        "vibe-check-mcp-server"
      ]
    }
  ]
}

Continue использует массив объектов серверов, а не map.

~/.config/zed/settings.json
{
  "context_servers": {
    "vibe-check": {
      "command": {
        "path": "npx",
        "args": [
          "-y",
          "vibe-check-mcp-server"
        ]
      }
    }
  }
}

Добавьте в context_servers. Zed перезагружается автоматически.

claude mcp add vibe-check -- npx -y vibe-check-mcp-server

Однострочная команда. Проверить: claude mcp list. Удалить: claude mcp remove.

Сценарии использования

Реальные сценарии: vibe-check-mcp-server

Sanity-check a plan before a destructive action with vibe-check

👤 Anyone running agents with write access ⏱ ~10 min beginner

Когда использовать: The agent is about to delete files, run a migration, or deploy — you want a second opinion.

Поток
  1. Invoke vibe_check on the plan
    Before running the migration, call vibe_check on this plan: [paste]. What could go wrong?✓ Скопировано
    → Pushback list: unchecked assumptions, missing rollback, scope creep
  2. Resolve or ignore each concern explicitly
    Address each concern — either mitigate or explicitly dismiss as out of scope.✓ Скопировано
    → Annotated plan ready to run

Итог: Fewer 'oh no' moments after destructive actions.

Подводные камни
  • Agent dismisses all concerns rubber-stamp style — Require the agent to restate each concern and the specific mitigation
Сочетать с: sequentialthinking-tools

Enforce session rules with vibe-check's constitution

👤 Teams sharing agent workflows ⏱ ~10 min intermediate

Когда использовать: You want certain rules (never push to main, always write tests first) enforced across a session.

Поток
  1. Set the constitution
    Use update_constitution to set these rules for this session: [list]. Enforce via CPI on every tool call that could violate them.✓ Скопировано
    → Rules acknowledged; subsequent violations get interrupted
  2. Inspect when in doubt
    Call check_constitution to see what's currently in effect.✓ Скопировано
    → Current rules list

Итог: Consistent agent behavior within a session without repeating rules every message.

Capture lessons from agent mistakes with vibe_learn

👤 Agent developers iterating on prompts ⏱ ~5 min beginner

Когда использовать: The agent made a classic mistake; you want it not to repeat.

Поток
  1. Log the lesson
    Call vibe_learn with: mistake='deleted prod table without snapshot', lesson='always take a snapshot before DDL on prod'.✓ Скопировано
    → Lesson stored for the session
  2. Rely on it later
    (later in same session) Attempt a DDL — vibe_check surfaces the stored lesson✓ Скопировано
    → Interrupt with the relevant past lesson

Итог: In-session memory of 'don't do that again' without prompt engineering.

Комбинации

Сочетайте с другими MCP — эффект x10

Plan + pushback loop

Plan with sequentialthinking_tools, then call vibe_check on the plan before executing step 1.✓ Скопировано
vibe-check + overture

Visual plan + vibe-check layer

After Overture produces the plan graph, run vibe_check on it and surface concerns as annotations.✓ Скопировано

Инструменты

Что предоставляет этот MCP

ИнструментВходные данныеКогда вызыватьСтоимость
vibe_check plan: str, context?: str Before any non-trivial or irreversible action 1 LLM call via your configured provider
vibe_learn mistake: str, lesson: str Immediately after a mistake, while fresh 0
update_constitution rules: str[] Session start or policy change 0
check_constitution Debug unexpected interrupts 0
reset_constitution Start fresh 0

Стоимость и лимиты

Во что обходится

Квота API
Uses your chosen LLM's quota
Токенов на вызов
500-3000 for vibe_check
Деньги
LLM token cost only
Совет
Only vibe_check on substantive plans — skipping it on trivial edits saves tokens

Безопасность

Права, секреты, радиус поражения

Хранение учётных данных: LLM API key in env (OPENAI_API_KEY, ANTHROPIC_API_KEY, etc.)
Исходящий трафик: Only to your configured LLM provider

Устранение неполадок

Частые ошибки и исправления

vibe_check returns empty concerns

Plan input was too vague — include concrete steps and side effects

Agent ignores vibe_check output

Make the system prompt require addressing each concern before proceeding

npx cannot find @pv-bhat/vibe-check-mcp

The published scope is @pv-bhat; pin latest explicitly: npx -y @pv-bhat/vibe-check-mcp@latest

Альтернативы

vibe-check-mcp-server в сравнении

АльтернативаКогда использоватьКомпромисс
sequentialthinking-toolsYou want planning, not pushbackDifferent shape — no adversarial second model
overtureYou want visual gating instead of text pushbackHeavier setup

Ещё

Ресурсы

📖 Читать официальный README на GitHub

🐙 Открытые задачи

🔍 Все 400+ MCP-серверов и Skills