Vibe Check MCP Server

What PV-Bhat/vibe-check-mcp-server MCP server does

PV-Bhat/vibe-check-mcp-server MCP server provides an oversight layer for AI agents. It is designed to interrupt flawed or overcomplicated plans before they continue, especially when the agent is drifting from the user's goal or approaching an irreversible action. The server asks a second large language model for metacognitive feedback rather than directly replacing the primary agent's reasoning.

The project focuses on two related problems: pattern inertia, where an agent continues a familiar approach despite warning signs, and reasoning lock-in, where a confident but flawed plan becomes difficult to change. Its feedback can include traits, uncertainty, and risk scores that help determine whether the agent should pause and reconsider.

How it works

An agent can call the Vibe Check functionality during a task, with system-prompt guidance recommending checks before irreversible operations. Chain-Pattern Interrupts provide phase-aware prompts that challenge assumptions. A separate intervention policy can consume the resulting signals and decide when the main agent should be interrupted before resuming work.

When a session ID is supplied, the server can summarize earlier advice to maintain continuity. The optional vibe_learn capability records mistakes and fixes for later reflection. The README also documents integration with the separate CPI project, including an architecture overview and an example handoff shim.

Setup and configuration

The documented quick-start path requires Node.js 20 or newer and the published npm package. For an MCP client using standard input/output, run:

npx -y @pv-bhat/vibe-check-mcp start --stdio

The same command can be added to an MCP client configuration with npx as the command and -y, @pv-bhat/vibe-check-mcp, start, and --stdio as arguments. The README specifically names Claude Desktop, Cursor, and Windsurf as compatible clients.

For local HTTP inspection, start the server with --http --port 2091. The health endpoint is /healthz, and JSON-RPC requests use /mcp. The server supports Gemini, Claude, GPT, and OpenRouter providers, but the supplied material does not specify provider credential names or a complete configuration schema.

Tools and capabilities

PV-Bhat/vibe-check-mcp-server MCP server supports:

• Metacognitive feedback from a second LLM.
• Adaptive, phase-aware Chain-Pattern Interrupts.
• Risk, uncertainty, and trait signals for oversight.
• Session history summaries when sessionId is provided.
• Optional learning records through vibe_learn.
• STDIO and streamable HTTP transports.

Limitations and notes

The project is in maintenance mode. Active feature development has ended, with future releases limited to maintenance patches, security fixes, and bug fixes; version 2.9.0 is identified as the latest maintenance release in the supplied material. Community forks and contributions are permitted under the MIT license.

The server depends on a second LLM for its review layer, so provider availability and configuration are relevant to deployment. The README describes supported model families and providers but does not establish that the service includes model access or specify whether provider usage incurs separate charges. Its documented role is advisory oversight: it supplies interruption and reflection signals rather than guaranteeing that an agent will never make an unsafe or misaligned decision.