prompt2mcp 0.1.0

Build MCP servers from natural language prompts — local-first, privacy-focused

MCP Factory 🏭

Build MCP servers from natural language prompts — local-first, privacy-focused

What is MCP Factory?

MCP Factory is a CLI tool that generates fully functional Model Context Protocol servers from plain English descriptions. Describe what you want, get production-ready TypeScript or Python code — complete with tool annotations, auth setup, and MCP best practices.

Installation

pip install mcp-factory

Or install from source:

git clone https://github.com/jatin/mcp-factory.git
cd mcp-factory
pip install -e ".[dev]"

Quick Start

# Create your first MCP server
mcpfactory create "Read my CSV files and answer questions about the data"

# Generate a GitHub API server (auto-detects API, generates auth + SETUP.md)
mcpfactory create "Create tools for the GitHub API" --name github-tools

# Python server with interactive refinement
mcpfactory create "Build a Slack bot" --lang python --interactive

# List your servers
mcpfactory list-servers

# See available templates
mcpfactory templates

# See supported APIs with auto-setup
mcpfactory supported-apis

Features

🎯 Prompt-to-Server

Describe what you want → get a working MCP server with tools, schemas, and error handling.

📦 8 Built-in Templates

Template	Description	Example Prompt
File Reader	Read, search, list local files	"Read and search through my project files"
Database Connector	Query SQL/NoSQL databases	"Connect to PostgreSQL and run queries"
API Wrapper	Wrap any REST API with tools	"Create tools for the GitHub API"
Web Scraper	Extract data from websites	"Scrape product prices from websites"
Document Processor	Parse PDFs, DOCX, Markdown	"Extract text from PDF documents"
Auth Server	JWT auth, user management, RBAC	"Build an authentication server with JWT"
Data Pipeline	ETL, transform, aggregate data	"ETL pipeline to ingest CSV and transform"
Notification Hub	Email, webhook, multi-channel alerts	"Send notifications via email and webhooks"

🤖 LLM-Powered Intelligence

MCP Factory uses a multi-tier generation pipeline:

1. LLM Prompt Analysis — Understands your intent using Ollama (local) or OpenAI/Claude (cloud)
2. API-Specific Tool Generation — Pre-built tool implementations for 5 APIs (GitHub, Slack, Stripe, Notion, Discord)
3. LLM Code Review — Auto-reviews generated code for quality, security, and MCP compliance
4. Interactive Mode — Asks follow-up questions for vague prompts to generate better servers
5. Keyword Fallback — Works fully offline without any LLM when needed

🔑 11 APIs with Auto-Setup

When your prompt mentions a supported API, MCP Factory automatically generates:

• .env.example — API-specific environment variables
• SETUP.md — Step-by-step guide to get API keys
• Auth validation — Startup checks with actionable error messages
• Pre-configured headers — Bearer, Basic, or OAuth2

API	Auth	Env Variable	Free Tier
GitHub	Bearer	GITHUB_TOKEN	✅
Slack	Bearer	SLACK_BOT_TOKEN	✅
OpenAI	Bearer	OPENAI_API_KEY	❌
Stripe	Bearer	STRIPE_SECRET_KEY	✅
Notion	Bearer	NOTION_API_KEY	✅
Spotify	OAuth2	SPOTIFY_CLIENT_ID	✅
Google	Bearer	GOOGLE_API_KEY	✅
Twitter/X	Bearer	TWITTER_BEARER_TOKEN	✅
Discord	Bearer	DISCORD_BOT_TOKEN	✅
Linear	Bearer	LINEAR_API_KEY	✅
Jira	Basic	JIRA_API_TOKEN	✅

💬 Interactive Mode

For vague prompts, MCP Factory asks targeted follow-up questions to generate better servers:

mcpfactory create "build something for files" --interactive

# MCP Factory asks:
# 1. What file types? (code, documents, data files, all)
# 2. What operations? (read, write, search, all)
# → Generates a refined, specific server

✅ MCP Best Practices

Generated servers follow MCP best practices:

• Tool annotations — readOnlyHint, destructiveHint, idempotentHint, openWorldHint
• Prefixed naming — Consistent tool prefixes (e.g., gh_, slack_)
• Schema descriptions — Zod .describe() / Pydantic Field(description=...) on every parameter
• Actionable errors — Errors include what went wrong + what to do next
• Pagination — List operations support cursor/maxResults parameters
• Secrets via env vars — Never hardcoded, always from .env

🌍 TypeScript & Python

Generate servers in your preferred language:

mcpfactory create "..." --lang typescript  # default
mcpfactory create "..." --lang python

🔒 Privacy First

• All generated code stays on your local filesystem
• Server metadata stored in local SQLite (~/.mcpfactory/servers.db)
• No telemetry, no cloud, no data collection
• LLM calls can use local Ollama models for complete offline privacy

CLI Reference

mcpfactory create <prompt>     Create a new MCP server
  --name, -n        Server name
  --lang, -l        Language: typescript | python (default: typescript)
  --output, -o      Output directory (default: ./output)
  --provider, -p    LLM provider: ollama | openai (default: ollama)
  --model, -m       Model name (e.g., llama3, gpt-4)
  --interactive      Enable follow-up questions for vague prompts
  --no-interactive   Skip interactive refinement

mcpfactory list-servers        List all generated servers
mcpfactory templates           Show available templates
mcpfactory supported-apis      Show APIs with auto-setup

Architecture

User Prompt
    │
    ├─► Interactive Refinement (follow-up questions if vague)
    │
    ├─► LLM Prompt Analyzer (Ollama/OpenAI/Claude)
    │       │
    │       └─► Keyword Fallback (offline)
    │
    ├─► API Detector (11 APIs with auth info)
    │
    ├─► Template Matcher (8 templates)
    │
    ├─► Code Generator
    │       ├─► Pre-built API tools (GitHub, Slack, Stripe, Notion, Discord)
    │       ├─► LLM tool generation
    │       └─► Template-based generation
    │
    ├─► Validator (syntax + structure checks)
    │
    ├─► LLM Code Reviewer (quality, security, MCP compliance)
    │
    └─► Local Storage (SQLite)

Development

# Install with dev dependencies
pip install -e ".[dev]"

# Run tests
pytest

# Lint
ruff check .

License

MIT