mcp-name: io.github.OtherVibes/mcp-as-a-judge

MCP as a Judge acts as a validation layer between AI coding assistants and LLMs, helping ensure safer and higher-quality code.

MCP as a Judge is a behavioral MCP that strengthens AI coding assistants by requiring explicit LLM evaluations for:

• Research, system design, and planning
• Code changes, testing, and task-completion verification

It enforces evidence-based research, reuse over reinvention, and human-in-the-loop decisions.

If your IDE has rules/agents (Copilot, Cursor, Claude Code), keep using them—this Judge adds enforceable approval gates on plan, code diffs, and tests.

• Treat LLM output as ground truth; skip research and use outdated information
• Reinvent the wheel instead of reusing libraries and existing code
• Cut corners: code below engineering standards and weak tests
• Make unilateral decisions when requirements are ambiguous or plans change
• Security blind spots: missing input validation, injection risks/attack vectors, least‑privilege violations, and weak defensive programming

• Evidence‑based research and reuse (best practices, libraries, existing code)
• Plan‑first delivery aligned to user requirements
• Human‑in‑the‑loop decisions for ambiguity and blockers
• Quality gates on code and tests (security, performance, maintainability)

• Intelligent code evaluation via MCP sampling; enforces software‑engineering standards and flags security/performance/maintainability risks
• Comprehensive plan/design review: validates architecture, research depth, requirements fit, and implementation approach
• User‑driven decisions via MCP elicitation: clarifies requirements, resolves obstacles, and keeps choices transparent
• Security validation in system design and code changes

MCP as a Judge is heavily dependent on MCP Sampling and MCP Elicitation features for its core functionality:

• MCP Sampling - Required for AI-powered code evaluation and judgment
• MCP Elicitation - Required for interactive user decision prompts

• Docker Desktop / Python 3.13+ - Required for running the MCP server

✅ Recommended setup: GitHub Copilot + VS Code — full MCP sampling; no API key needed.

⚠️ Critical: For assistants without full MCP sampling (Cursor, Claude Code, Augment, Qodo), you MUST set LLM_API_KEY. Without it, the server cannot evaluate plans or code. See LLM API Configuration.

💡 Tip: Prefer large context models (≥ 1M tokens) for better analysis and judgments.

For troubleshooting, visit the FAQs section.

Configure MCP as a Judge in your MCP-enabled client:

Notes:

• VS Code controls the sampling model; select it via “MCP: List Servers → mcp-as-a-judge → Configure Model Access”.

1. Configure MCP Settings:

   Add this to your MCP client configuration file:

   {
     "command": "docker",
     "args": ["run", "--rm", "-i", "--pull=always", "ghcr.io/othervibes/mcp-as-a-judge:latest"],
     "env": {
       "LLM_API_KEY": "your-openai-api-key-here",
       "LLM_MODEL_NAME": "gpt-4o-mini"
     }
   }

   📝 Configuration Options (All Optional):

   • LLM_API_KEY: Optional for GitHub Copilot + VS Code (has built-in MCP sampling)
   • LLM_MODEL_NAME: Optional custom model (see Supported LLM Providers for defaults)
   • The --pull=always flag ensures you always get the latest version automatically

   Then manually update when needed:

   # Pull the latest version
   docker pull ghcr.io/othervibes/mcp-as-a-judge:latest

1. Install the package:

   uv tool install mcp-as-a-judge

2. Configure MCP Settings:

   The MCP server may be automatically detected by your MCP‑enabled client.

   📝 Notes:

   • No additional configuration needed for GitHub Copilot + VS Code (has built-in MCP sampling)
   • LLM_API_KEY is optional and can be set via environment variable if needed

3. To update to the latest version:

   # Update MCP as a Judge to the latest version
   uv tool upgrade mcp-as-a-judge

• Open Command Palette (Cmd/Ctrl+Shift+P) → “MCP: List Servers”
• Select the configured server “mcp-as-a-judge”
• Choose “Configure Model Access”
• Check your preferred model(s) to enable sampling

For AI assistants without full MCP sampling support you can configure an LLM API key as a fallback. This ensures MCP as a Judge works even when the client doesn't support MCP sampling.

• Set LLM_API_KEY (unified key). Vendor is auto-detected; optionally set LLM_MODEL_NAME to override the default.

1. Open Cursor Settings:

   • Go to File → Preferences → Cursor Settings
   • Navigate to the MCP tab
   • Click + Add to add a new MCP server

2. Add MCP Server Configuration:

   {
     "command": "uv",
     "args": ["tool", "run", "mcp-as-a-judge"],
     "env": {
       "LLM_API_KEY": "your-openai-api-key-here",
       "LLM_MODEL_NAME": "gpt-4.1"
     }
   }

   📝 Configuration Options:

   • LLM_API_KEY: Required for Cursor (limited MCP sampling)
   • LLM_MODEL_NAME: Optional custom model (see Supported LLM Providers for defaults)

1. Add MCP Server via CLI:

   # Set environment variables first (optional model override)
   export LLM_API_KEY="your_api_key_here"
   export LLM_MODEL_NAME="claude-3-5-haiku"  # Optional: faster/cheaper model
   
   # Add MCP server
   claude mcp add mcp-as-a-judge -- uv tool run mcp-as-a-judge

2. Alternative: Manual Configuration:

   • Create or edit ~/.config/claude-code/mcp_servers.json

   {
     "command": "uv",
     "args": ["tool", "run", "mcp-as-a-judge"],
     "env": {
       "LLM_API_KEY": "your-anthropic-api-key-here",
       "LLM_MODEL_NAME": "claude-3-5-haiku"
     }
   }

   📝 Configuration Options:

   • LLM_API_KEY: Required for Claude Code (limited MCP sampling)
   • LLM_MODEL_NAME: Optional custom model (see Supported LLM Providers for defaults)

For other MCP-compatible clients, use the standard MCP server configuration:

{
  "command": "uv",
  "args": ["tool", "run", "mcp-as-a-judge"],
  "env": {
    "LLM_API_KEY": "your-openai-api-key-here",
    "LLM_MODEL_NAME": "gpt-5"
  }
}

📝 Configuration Options:

• LLM_API_KEY: Required for most MCP clients (except GitHub Copilot + VS Code)
• LLM_MODEL_NAME: Optional custom model (see Supported LLM Providers for defaults)

Primary Mode: MCP Sampling

• All judgments are performed using MCP Sampling capability
• No need to configure or pay for external LLM API services
• Works directly with your MCP-compatible client's existing AI model
• Currently supported by: GitHub Copilot + VS Code

Fallback Mode: LLM API Key

• When MCP sampling is not available, the server can use LLM API keys
• Supports multiple providers via LiteLLM: OpenAI, Anthropic, Google, Azure, Groq, Mistral, xAI
• Automatic vendor detection from API key patterns
• Default model selection per vendor when no model is specified

• The server runs locally on your machine
• No data collection - your code and conversations stay private
• No external API calls when using MCP Sampling. If you set LLM_API_KEY for fallback, the server will call your chosen LLM provider only to perform judgments (plan/code/test) with the evaluation content you provide.
• Complete control over your development workflow and sensitive information

We welcome contributions! Please see CONTRIBUTING.md for guidelines.

# Clone the repository
git clone https://github.com/OtherVibes/mcp-as-a-judge.git
cd mcp-as-a-judge

# Install dependencies with uv
uv sync --all-extras --dev

# Install pre-commit hooks
uv run pre-commit install

# Run tests
uv run pytest

# Run all checks
uv run pytest && uv run ruff check && uv run ruff format --check && uv run mypy src

© 2025 OtherVibes and Zvi Fried. The "MCP as a Judge" concept, the "behavioral MCP" approach, the staged workflow (plan → code → test → completion), tool taxonomy/descriptions, and prompt templates are original work developed in this repository.

While “LLM‑as‑a‑judge” is a broadly known idea, this repository defines the original “MCP as a Judge” behavioral MCP pattern by OtherVibes and Zvi Fried. It combines task‑centric workflow enforcement (plan → code → test → completion), explicit LLM‑based validations, and human‑in‑the‑loop elicitation, along with the prompt templates and tool taxonomy provided here. Please attribute as: “OtherVibes – MCP as a Judge (Zvi Fried)”.

• References: GitHub Copilot Custom Instructions, Cursor Rules, Claude Code Subagents

• Tasklist = planning/organization: tracks tasks, priorities, and status. It doesn’t guarantee engineering quality or readiness.
• Judge workflow = quality gates: enforces approvals for plan/design, code diffs, tests, and final completion. It demands real evidence (e.g., unified Git diffs and raw test output) and returns structured approvals and required improvements.
• Together: Use the tasklist to organize work; use the Judge to decide when each stage is actually ready to proceed. The server also emits next_tool guidance to keep progress moving through the gates.

• In your prompt: "use mcp-as-a-judge" or "Evaluate plan/code/test using the MCP server mcp-as-a-judge".
• VS Code: Command Palette → "MCP: List Servers" → ensure "mcp-as-a-judge" is listed and enabled.
• Ensure the MCP server is running and, in your client, the judge tools are enabled/approved.

• Open Command Palette (Cmd/Ctrl+Shift+P) → "MCP: List Servers"
• Select "mcp-as-a-judge" → "Configure Model Access"
• Check your preferred model(s) to enable sampling

This project is licensed under the MIT License (see LICENSE).

• Model Context Protocol by Anthropic
• LiteLLM for unified LLM API integration