Vibe Check

Vibe Check

Plug-and-play mentor layer for MCP servers to prevent overengineering and align LLMs.

322
Stars
41
Forks
7
Releases

Overview

Vibe Check MCP is a lightweight MCP server that acts as an external metacognitive mentor for AI agents. It implements Anthropic's Model Context Protocol and couples a vibe-check signal layer with Chain-Pattern Interrupts (CPI) to pause at risk inflection moments and re-align the agent’s actions before irreversible steps. The server surfaces traits, uncertainty, and risk scores, while a secondary LLM provides meta-cognitive feedback to guide the agent’s next moves. By enforcing per-session constitutions and an optional vibe_learn history, it reduces pattern inertia, misalignment, and overengineering, improving reliability and safety. It supports multiple providers (Gemini, OpenAI, OpenRouter, Anthropic) and offers transport options via STDIO or HTTP, with quickstart commands using npx. The architecture, CPI integration guide, and client docs are included, along with references to CPI research and MCP registries. It includes per-session APIs (update_constitution, reset_constitution, check_constitution) and a plug-and-play approach compatible with common MCP clients like Claude Desktop, Cursor, Windsurf, and VS Code.

Details

Owner
PV-Bhat
Language
TypeScript
License
MIT License
Updated
2025-12-07

Features

CPI Adaptive Interrupts

Phase-aware prompts that challenge assumptions, providing alignment and robustness by interrupting risky reasoning steps at inflection points.

Multi-provider LLM

Supports Gemini, OpenAI, Anthropic, and OpenRouter, offering flexibility in provider choice and integration.

History Continuity

Summarizes prior advice when a sessionId is supplied to maintain context across interactions.

Optional vibe_learn

Logs mistakes and fixes for future reflection, enabling self-improvement over time.

Session Constitution (per-session rules)

API set to update, reset, and check session-specific rules to enforce CPI-guided behavior.

Transport & Client Plug-and-Play

Supports stdio and HTTP transports with quickstart npx commands and compatibility with common MCP clients (Claude Desktop, Cursor, Windsurf, VS Code).

Audience

AI developersUse as an external mentor layer to keep agents aligned, reflective, and safe via CPI.
MCP integratorsIntegrate across MCP transports and clients for oversight and minimal viable-agent guidance.
ResearchersStudy CPI-based mentorship and RLI mitigation as part of alignment research.

Tags

MCPVibe Check MCPCPIChain-Pattern Interruptmetacognitive mentorLLM alignmentReasoning Lock-Insafetymulti-provider LLM