Meta Prompter MCP

A tiny meta‑prompt MCP server that grades prompts and returns JSON‑only.

What it does

Evaluates the prompt you pass in as tool argument across 8 dimensions (clarity, specificity, context, actionability, safety, testability, hallucination prevention, token efficiency), computes a weighted global score, and—if needed—returns a full rewrite. Zero fluff, machine‑friendly.

Quick theory

• On the basis of the paper Meta-Prompting

• The model’s task is to analyze another prompt, not perform the end‑task — prompts about prompts = meta.

• LLM‑as‑a‑Judge: role + rubric + constrained JSON output approximates structured human evaluation.

• Form‑filling graders (e.g., G‑Eval): schema‑locked fields and brief justifications reduce variance and make results comparable.

• Rule‑guided critique (Constitutional‑style): conflict rules, edge‑case handling, and safety checks function as a small “constitution.”

• Reflection/self‑critique patterns: institutionalizes a critique step that boosts reliability before you ship a prompt.

One‑liner: it treats prompts as artifacts to be graded with rules, not instructions to be executed.

Why MCP

Technical Foundation: Claude Code doesn't allow customizing the LLM temperature. Its default temperature is 1. For more stable scoring, use a lower temperature — this reduces randomness in sampling, ensuring that scores and justifications remain consistent across runs. The temperature is 0 in MCP server.

Agentic Integration: MCP enables sophisticated agent workflows by providing machine-readable evaluation data that agents can act upon programmatically:

Configure MCP

1. ./setup.sh
2. claude mcp add meta-prompter --env PROMPT_EVAL_MODEL=sonnet-4 --env PROMPT_EVAL_API_KEY=<claude_api_key> -- npx -y meta-prompter-mcp@latest

JSON Configuration

    "meta-prompter": {
      "command": "npx",
      "args": [
        "-y",
        "meta-prompter-mcp@latest"
      ],
      "env": {
        "PROMPT_EVAL_MODEL": "sonnet-4",
        "PROMPT_EVAL_API_KEY": "sk-123456789",
        "ANTHROPIC_BASE_URL": "http://localhost:4000/anthropic/v1"
      }
    }

Environment Variables:

• PROMPT_EVAL_MODEL - Model name (see mappings below)
• PROMPT_EVAL_API_KEY - API key
• OPENAI_BASE_URL - Optional OpenAI endpoint
• ANTHROPIC_BASE_URL - Optional Anthropic endpoint

Model Name Mappings

The following short names map to their full model identifiers:

Anthropic Models: | Short Name | Full Model ID | |-----------|---------------| | opus-4-1 | claude-opus-4-1-20250805 | | sonnet-4 | claude-sonnet-4-20250514 | | sonnet-4.5 | claude-sonnet-4-5-20250929 |

OpenAI Models: | Short Name | Full Model ID | |-----------|---------------| | gpt-5 | gpt-5 |

Available Tools

• evaluate - Evaluate a prompt using AI analysis.
  • prompt (string, required): The prompt to evaluate
• ping - Simple ping test to verify connection

Available Resources

• file:///eval-viewer.html - SPA for viewing evaluation results

View Eval Results

The MCP tool result json line created per command run is appended to evaluation_result.jsonl in current claude code working dir.

eval-viewer.html is the SPA used to view the eval result jsonl data.

Open the viewer in browser and upload the evaluation_result.jsonl to view the result such as

Testing

# Test with MCP Inspector
npx @modelcontextprotocol/inspector node dist/index.js

Publishing

• Ensure the latest version is committed in the package.json following Semantic Versioning. e.g. "version": "1.1.1"
• Run ./setup.sh
• Run npm publish --tag latest