Promptheus MCP Server

Promptheus

Refine and optimize prompts for LLMs

<!-- mcp-name: io.github.abhichandra21/promptheus -->

Quick Start

pip install promptheus

# Interactive session
promptheus

# Single prompt
promptheus "Write a technical blog post"

# Skip clarifying questions
promptheus -s "Explain Kubernetes"

# Use web UI
promptheus web

Python library usage

from promptheus import refine_prompt

result = refine_prompt("Write a technical blog post", skip_questions=True)
print(result["refined_prompt"])

If you're already in an async application (e.g., FastAPI), call refine_prompt_async instead of the sync helper.

What is Promptheus?

Promptheus analyzes your prompts and refines them with:

• Adaptive questioning: Smart detection of what information you need to provide
• Multi-provider support: Works with Google, OpenAI, Anthropic, Groq, Qwen, and more
• Interactive refinement: Iteratively improve outputs through natural conversation
• Session history: Automatically track and reuse past prompts
• CLI and Web UI: Use from terminal or browser

Supported Providers

Provider	Models	Setup
Google Gemini	gemini-2.0-flash, gemini-1.5-pro	API Key
Anthropic Claude	claude-3-5-sonnet, claude-3-opus	Console
OpenAI	gpt-4o, gpt-4-turbo	API Key
Groq	llama-3.3-70b, mixtral-8x7b	Console
Alibaba Qwen	qwen-max, qwen-plus	DashScope
Zhipu GLM	glm-4-plus, glm-4-air	Console
OpenRouter	openrouter/auto (auto-routing)	Dashboard

OpenRouter integration in Promptheus is optimized around the openrouter/auto routing model:

• Model listing is intentionally minimal: Promptheus does not expose your full OpenRouter account catalog.
• You can still specify a concrete model manually with OPENROUTER_MODEL or --model if your key has access.

Core Features

🧠 Adaptive Task Detection Automatically detects whether your task needs refinement or direct optimization

⚡ Interactive Refinement Ask targeted questions to elicit requirements and improve outputs

📝 Pipeline Integration Works seamlessly in Unix pipelines and shell scripts

🔄 Session Management Track, load, and reuse past prompts automatically

📊 Telemetry & Analytics Anonymous usage and performance metrics tracking for insights (local storage only, can be disabled)

🌐 Web Interface Beautiful UI for interactive prompt refinement and history management

Configuration

Create a .env file with at least one provider API key:

GOOGLE_API_KEY=your_key_here
ANTHROPIC_API_KEY=your_key_here
OPENAI_API_KEY=your_key_here

Or run the interactive setup:

promptheus auth

Examples

Content Generation

promptheus "Write a blog post about async programming"
# System asks: audience, tone, length, key topics
# Generates refined prompt with all specifications

Code Analysis

promptheus -s "Review this function for security issues"
# Skips questions, applies direct enhancement

Interactive Session

promptheus
/set provider anthropic
/set model claude-3-5-sonnet
# Process multiple prompts, switch providers/models with /commands

Pipeline Integration

echo "Create a REST API schema" | promptheus | jq '.refined_prompt'
cat prompts.txt | while read line; do promptheus "$line"; done

Testing & Examples: See sample_prompts.md for test prompts demonstrating adaptive task detection (analysis vs generation).

Telemetry & Analytics

# View telemetry summary (anonymous metrics about usage and performance)
promptheus telemetry summary

# Disable telemetry if desired
export PROMPTHEUS_TELEMETRY_ENABLED=0

# Customize history storage location
export PROMPTHEUS_HISTORY_DIR=~/.custom_promptheus

MCP Server

Promptheus includes a Model Context Protocol (MCP) server that exposes prompt refinement capabilities as standardized tools for integration with MCP-compatible clients.

What the MCP Server Does

The Promptheus MCP server provides:

• Prompt refinement with Q&A: Intelligent prompt optimization through adaptive questioning
• Prompt tweaking: Surgical modifications to existing prompts
• Model/provider inspection: Discovery and validation of available AI providers
• Environment validation: Configuration checking and connectivity testing

Starting the MCP Server

# Start the MCP server
promptheus mcp

# Or run directly with Python
python -m promptheus.mcp_server

Prerequisites:

• MCP package installed: pip install mcp (included in requirements.txt)
• At least one provider API key configured (see Configuration)

Available MCP Tools

refine_prompt

Intelligent prompt refinement with optional clarification questions.

Inputs:

• prompt (required): The initial prompt to refine
• answers (optional): Dictionary mapping question IDs to answers {q0: "answer", q1: "answer"}
• answer_mapping (optional): Maps question IDs to original question text
• provider (optional): Override provider (e.g., "google", "openai")
• model (optional): Override model name

Response Types:

• {"type": "refined", "prompt": "...", "next_action": "..."}: Success with refined prompt
• {"type": "clarification_needed", "questions_for_ask_user_question": [...], "answer_mapping": {...}}: Questions needed
• {"type": "error", "error_type": "...", "message": "..."}: Error occurred

tweak_prompt

Apply targeted modifications to existing prompts.

Inputs:

• prompt (required): Current prompt to modify
• modification (required): Description of changes (e.g., "make it shorter")
• provider, model (optional): Provider/model overrides

Returns:

• {"type": "refined", "prompt": "..."}: Modified prompt

list_models

Discover available models from configured providers.

Inputs:

• providers (optional): List of provider names to query
• limit (optional): Max models per provider (default: 20)
• include_nontext (optional): Include vision/embedding models

Returns:

• {"type": "success", "providers": {"google": {"available": true, "models": [...]}}}

list_providers

Check provider configuration status.

Returns:

• {"type": "success", "providers": {"google": {"configured": true, "model": "..."}}}

validate_environment

Test environment configuration and API connectivity.

Inputs:

• providers (optional): Specific providers to validate
• test_connection (optional): Test actual API connectivity

Returns:

• {"type": "success", "validation": {"google": {"configured": true, "connection_test": "passed"}}}

Prompt Refinement Workflow with Q&A

The MCP server supports a structured clarification workflow for optimal prompt refinement:

Step 1: Initial Refinement Request

{
  "tool": "refine_prompt",
  "arguments": {
    "prompt": "Write a blog post about machine learning"
  }
}

Step 2: Handle Clarification Response

{
  "type": "clarification_needed",
  "task_type": "generation",
  "message": "To refine this prompt effectively, I need to ask...",
  "questions_for_ask_user_question": [
    {
      "question": "Who is your target audience?",
      "header": "Q1",
      "multiSelect": false,
      "options": [
        {"label": "Technical professionals", "description": "Technical professionals"},
        {"label": "Business executives", "description": "Business executives"}
      ]
    }
  ],
  "answer_mapping": {
    "q0": "Who is your target audience?"
  }
}

Step 3: Collect User Answers

Use your MCP client's AskUserQuestion tool with the provided questions, then map answers to question IDs.

Step 4: Final Refinement with Answers

{
  "tool": "refine_prompt", 
  "arguments": {
    "prompt": "Write a blog post about machine learning",
    "answers": {"q0": "Technical professionals"},
    "answer_mapping": {"q0": "Who is your target audience?"}
  }
}

Response:

{
  "type": "refined",
  "prompt": "Write a comprehensive technical blog post about machine learning fundamentals targeted at software engineers and technical professionals. Include practical code examples and architectural patterns...",
  "next_action": "This refined prompt is now ready to use. If the user asked you to execute/run the prompt, use this refined prompt directly with your own capabilities..."
}

AskUser Integration Contract

The MCP server operates in two modes:

Interactive Mode (when AskUserQuestion is available):

• Automatically asks clarification questions via injected AskUserQuestion function
• Returns refined prompt immediately after collecting answers
• Seamless user experience within supported clients

Structured Mode (fallback for all clients):

• Returns clarification_needed response with formatted questions
• Client responsible for calling AskUserQuestion tool
• Answers mapped back via answer_mapping dictionary

Question Format: Each question in questions_for_ask_user_question includes:

• question: The question text to display
• header: Short identifier (Q1, Q2, etc.)
• multiSelect: Boolean for multi-select options
• options: Array of {label, description} for radio/checkbox questions

Answer Mapping:

• Question IDs follow pattern: q0, q1, q2, etc.
• Answers dictionary uses these IDs as keys: {"q0": "answer", "q1": "answer"}
• answer_mapping preserves original question text for provider context

Troubleshooting MCP

MCP Package Not Installed

Error: The 'mcp' package is not installed. Please install it with 'pip install mcp'.

Fix: pip install mcp or install Promptheus with dev dependencies: pip install -e .[dev]

Missing Provider API Keys

{
  "type": "error",
  "error_type": "ConfigurationError", 
  "message": "No provider configured. Please set API keys in environment."
}

Diagnosis: Use list_providers or validate_environment tools to check configuration status

Provider Misconfiguration

{
  "type": "success",
  "providers": {
    "google": {"configured": false, "error": "GOOGLE_API_KEY not found"},
    "openai": {"configured": true, "model": "gpt-4o"}
  }
}

Fix: Set missing API keys in .env file or environment variables

Connection Test Failures

{
  "type": "success", 
  "validation": {
    "google": {
      "configured": true,
      "connection_test": "failed: Authentication error"
    }
  }
}

Fix: Verify API keys are valid and have necessary permissions

Full Documentation

Quick reference: promptheus --help

Comprehensive guides:

• 📖 Installation & Setup
• 🚀 Usage Guide
• 🔧 Configuration
• ⌨️ CLI Reference
• 🌐 Web UI Guide
• 🔌 Provider Setup

Development

git clone https://github.com/abhichandra21/Promptheus.git
cd Promptheus
pip install -e ".[dev]"
pytest -q

See CLAUDE.md for detailed development guidance.

License

MIT License - see LICENSE for details

Contributing

Contributions welcome! Please see our development guide for contribution guidelines.

---

Questions? Open an issue | Live demo: promptheus web