Vibe Check MCP Server for Claude

Safety notes

Vibe Check MCP is in final maintenance mode according to the project README, so check the repository and package state before adopting it broadly.
The server is an advisory oversight layer; it can challenge assumptions but cannot guarantee safe, correct, or complete decisions.
The `vibe_check` tool sends the goal, plan, optional user prompt, progress, uncertainties, and task context to a configured LLM provider for critique.
The `vibe_learn` tool can record mistakes, preferences, successes, categories, and solutions for later reflection.
Constitution tools can help apply session rules, but they do not enforce external system permissions or prevent client-side actions by themselves.

Privacy notes

Agent plans, prompts, progress, uncertainties, session identifiers, mistakes, and learned preferences can reveal confidential project details.
Provider API keys are loaded from environment variables and should be scoped, stored, and rotated according to local policy.
Depending on provider selection, request metadata and prompt content may be processed by Gemini, OpenAI, OpenRouter, Anthropic, or compatible endpoints.
Avoid sending secrets, private code, customer data, or regulated information in `vibe_check` or `vibe_learn` inputs unless approved for the selected provider.

Prerequisites

Node.js 20 or newer and npx available to the MCP client runtime.
At least one supported LLM provider credential, such as Gemini, OpenAI, OpenRouter, or Anthropic.
Review of the project's maintenance notice before relying on it for long-lived workflows.
Clear policy for what agent plans, prompts, progress notes, and session context may be sent to a second model provider.

Schema details

Install type: cli
Troubleshooting: No

Source repository stats

Scope: Source repo

Collection metadata

Estimated setup: 15 minutes
Difficulty: intermediate

Full copyable content

{
  "mcpServers": {
    "vibe-check-mcp": {
      "command": "npx",
      "args": ["-y", "@pv-bhat/vibe-check-mcp", "start", "--stdio"],
      "env": {
        "GEMINI_API_KEY": ""
      }
    }
  }
}

About this resource

Content

Vibe Check MCP is a maintenance-mode Model Context Protocol server for adding mentor-style plan review to agent workflows. It gives Claude tools for asking a second model to critique a goal and plan, recording recurring lessons, and checking session-specific rules before continuing complex work.

The server is designed for moments where an agent may be over-engineering, following an unsupported assumption, or preparing to take an irreversible step. It does not replace human review, tests, access control, or policy enforcement.

Source Review

These sources were reviewed on 2026-06-06. Prefer the live repository, README, npm registry metadata, package metadata, MCP registry manifest, source implementation, and license file for current install commands, maintenance status, supported providers, tool schemas, transport options, and licensing.

Features

npm package @pv-bhat/vibe-check-mcp.
Stdio MCP server launched with npx -y @pv-bhat/vibe-check-mcp start --stdio.
Optional HTTP transport documented upstream for clients that support it.
vibe_check tool for goal and plan critique with optional prompt, progress, uncertainty, task context, session ID, and provider override fields.
vibe_learn tool for recording mistakes, preferences, successes, categories, and solutions.
update_constitution, reset_constitution, and check_constitution tools for session rules.
Gemini, OpenAI, OpenRouter, Anthropic, and Anthropic-compatible endpoint support in the source implementation.
MIT license.

Installation

Configure the stdio server in your MCP client:

{
  "mcpServers": {
    "vibe-check-mcp": {
      "command": "npx",
      "args": ["-y", "@pv-bhat/vibe-check-mcp", "start", "--stdio"],
      "env": {
        "GEMINI_API_KEY": ""
      }
    }
  }
}

After restarting the MCP client, ask Claude to call vibe_check before a high-risk, ambiguous, or irreversible plan, then use the critique to revise the next step.

Use Cases

Challenge a coding agent's plan before a large refactor.
Ask for an independent critique before running destructive commands.
Record recurring mistakes and fixes with vibe_learn.
Attach temporary session rules to a long agent workflow.
Add a lightweight reflection step before ambiguous product or architecture work.
Compare a proposed plan against uncertainties and task context.

Safety and Privacy

Vibe Check MCP should be treated as advisory. A second model can provide useful critique, but it can also misunderstand context, miss risks, or produce overconfident guidance. Continue to rely on tests, review, access controls, and human judgment for important changes.

Inputs to vibe_check and vibe_learn may include private plans, prompts, progress notes, code context, lessons learned, and session identifiers. Those values can be sent to the configured model provider, so keep secrets, customer data, regulated information, and unreleased project details out of requests unless that provider and environment are approved.

Duplicate Check

Existing MCP content includes planning, coding, browser, and automation servers, but no dedicated entry for PV-Bhat/vibe-check-mcp-server or the @pv-bhat/vibe-check-mcp package was found in content/mcp. This entry is distinct because it covers metacognitive agent oversight tools rather than a general coding, browser, or task-management MCP server.

#agent-oversight#reflection#ai-safety#prompt-engineering#planning