The definitive Vibe Coder's sanity check MCP server: Prevent cascading errors in AI workflows by implementing strategic pattern interrupts. Uses tool call "Vibe Check" with LearnLM 1.5 Pro (Gemini API), fine-tuned for pedagogy and metacognition to enhance complex workflow strategy, and prevents tunnel vision errors.
Also find Vibecheck on: mcpservers.org, Glama.ai, mcp.so
Your AI’s inner rubber duck when it can’t rubber duck itself.
In the “vibe coding” era, AI agents now have incredible capabilities, but the question has now moved:
from
“Can my AI agent really do this complex task?”
to
“Can my AI agent understand that I want to write a simple program, not an infrastructure for a multi-billion dollar tech company?”
It provides the essential “Hold up… this ain’t it” moment that AI agents don’t currently have: a built in self-correcting oversight layer. It’s the definitive Vibe Coder’s sanity check MCP server:
TLDR; Implement an agent fine-tuned to stop your agent and make it reconsider before it confidently implements something wrong.
In the vibe coding movement, we’re all using LLMs to generate, refactor, and debug our code. But these models have a critical flaw: once they start down a reasoning path, they’ll keep going even when the path is clearly wrong.
You: "Parse this CSV file"
AI: "First, let's implement a custom lexer/parser combination that can handle arbitrary
CSV dialects with an extensible architecture for future file formats..."
You: *stares at 200 lines of code when you just needed to read 10 rows*
This pattern inertia leads to:
Vibe Check adds a metacognitive layer to your agent workflows with three integrated tools:
Pattern interrupt mechanism that breaks tunnel vision with metacognitive questioning:
vibe_check({
"phase": "planning", // planning, implementation, or review
"userRequest": "...", // FULL original user request
"plan": "...", // Current plan or thinking
"confidence": 0.7 // Optional: 0-1 confidence level
})
Meta-thinking anchor point that recalibrates complex workflows:
vibe_distill({
"plan": "...", // Detailed plan to simplify
"userRequest": "..." // FULL original user request
})
Self-improving feedback loop that builds pattern recognition over time:
vibe_learn({
"mistake": "...", // One-sentence description of mistake
"category": "...", // From standard categories
"solution": "..." // How it was corrected
})
Before Vibe Check:
Claude assumes the meaning of MCP despite ambiguity, leading to all subsequent steps having this wrong assumption
After Vibe Check:
Vibe Check MCP is called, and points out the ambiguity, which forces Claude to acknowledge this lack of information and proactively address it
To install vibe-check-mcp-server for Claude Desktop automatically via Smithery:
npx -y @smithery/cli install @PV-Bhat/vibe-check-mcp-server --client claude
## Clone the repo
git clone https://github.com/PV-Bhat/vibe-check-mcp-server.git
cd vibe-check-mcp-server
## Install dependencies
npm install
## Build the project
npm run build
## Start the server
npm run start
Add to your claude_desktop_config.json:
"vibe-check": {
"command": "node",
"args": [
"/path/to/vibe-check-mcp/build/index.js"
],
"env": {
"GEMINI_API_KEY": "YOUR_GEMINI_API_KEY"
}
}
Create a .env file in the project root:
GEMINI_API_KEY=your_gemini_api_key_here
For effective pattern interrupts, include these instructions in your system prompt:
As an autonomous agent, you will:
1. Treat vibe_check as a critical pattern interrupt mechanism
2. ALWAYS include the complete user request with each call
3. Specify the current phase (planning/implementation/review)
4. Use vibe_distill as a recalibration anchor when complexity increases
5. Build the feedback loop with vibe_learn to record resolved issues
| Tool | When to Use |
|---|---|
| 🛑 vibe_check | When your agent starts explaining blockchain fundamentals for a todo app |
| ⚓ vibe_distill | When your agent’s plan has more nested bullet points than your entire tech spec |
| 🔄 vibe_learn | After you’ve manually steered your agent back from the complexity abyss |
See the Technical Reference{:target=“_blank”} for complete API documentation.
Vibe Check implements a dual-layer metacognitive architecture based on recursive oversight principles. Key insights:
Pattern Inertia Resistance: LLM agents naturally demonstrate a momentum-like property in their reasoning paths, requiring external intervention to redirect.
Phase-Resonant Interrupts: Metacognitive questioning must align with the agent’s current phase (planning/implementation/review) to achieve maximum corrective impact.
Authority Structure Integration: Agents must be explicitly prompted to treat external metacognitive feedback as high-priority interrupts rather than optional suggestions.
Anchor Compression Mechanisms: Complex reasoning flows must be distilled into minimal anchor chains to serve as effective recalibration points.
Recursive Feedback Loops: All observed missteps must be stored and leveraged to build longitudinal failure models that improve interrupt efficacy.
For more details on the underlying design principles, see Philosophy{:target=“_blank”}.
| Document | Description |
|---|---|
| Agent Prompting Strategies{:target=“_blank”} | Detailed techniques for agent integration |
| Advanced Integration{:target=“_blank”} | Feedback chaining, confidence levels, and more |
| Technical Reference{:target=“_blank”} | Complete API documentation |
| Philosophy{:target=“_blank”} | The deeper AI alignment principles behind Vibe Check |
| Case Studies{:target=“_blank”} | Real-world examples of Vibe Check in action |
We welcome contributions to Vibe Check! Whether it’s bug fixes, feature additions, or just improving documentation, check out our Contributing Guidelines{:target=“_blank”} to get started.
MIT{:target=“_blank”}
