Enable AI agents to build and execute Pipelex pipelines on-the-fly. Just describe what you need, and watch as the agent constructs and runs the pipeline for you.

📑 Table of Contents

• Introduction

• Key Features

• Quick Start

  • Prerequisites

  • Installation

  • API Key Configuration

• Client Setup

  • Cursor

  • Claude Desktop

  • Other Clients

• Examples

• Known Limitations

• Contributing

• Support

• License

Introduction

👉 New to Pipelex? Check out the main repository to learn more!

The Pipelex MCP Server implements the Model Context Protocol to enable AI agents to interact with Pipelex as a native tool. This bridges the gap between conversational AI and structured, repeatable AI workflows.

With this MCP server, AI agents can:

• Build pipelines from natural language descriptions

• Execute pipelines with specific inputs

• Receive structured outputs ready for downstream tasks

To learn more about MCP, check out the official MCP documentation from Anthropic.

🌟 Key Features

🎯 Natural Language Pipeline Creation

Simply tell your AI agent what you need:

"Build me a pipeline that extracts the buyer's name from a purchase receipt, then validates the email format."

The agent will construct the pipeline definition and execute it with your inputs—all in one conversation.

🔧 Two Powerful Tools

1. pipe_builder: Constructs pipeline definitions from natural language

2. pipe_runner: Executes pipelines with structured inputs

🚀 Works with Any MCP Client

• Cursor

• Claude Desktop

• Any MCP-compatible client (not tested yet)

🚀 Quick Start

Prerequisites

• Python >=3.11,<3.12

• uv package manager (required)

⚠️ Important: The Pipelex MCP Server requires uv to be installed on your system. Make sure to install it before proceeding.

Installing uv

If you don't have uv installed:

# macOS/Linux
curl -LsSf https://astral.sh/uv/install.sh | sh

# Windows
powershell -c "irm https://astral.sh/uv/install.ps1 | iex"

# Verify installation
uv --version

After installation, get the full path to uv (you'll need this for MCP configuration):

which uv

Keep this path handy—you'll use it in your MCP client configuration.

Related MCP server: Devpipe MCP Server

Installation

1. Clone the repository:

git clone https://github.com/Pipelex/pipelex-mcp.git
cd pipelex-mcp

2. Install dependencies:

make install

API Key Configuration

Get Your Pipelex Inference API Key

The MCP server requires a Pipelex Inference API key to execute pipelines.

Get a free API key ($20 in free credits):

1. Join our Discord community

2. Request your API key in the 🔑・free-api-key channel

Set Up Environment Variables

# Copy the example environment file
cp server/.env.example server/.env

# Edit server/.env and add your API key
# PIPELEX_INFERENCE_API_KEY=your-api-key-here

Note: For advanced configuration (bring your own API keys, custom backends), see the Pipelex API Key Configuration guide.

🔌 Client Setup

Cursor

Cursor has built-in MCP support. The configuration file is located at .cursor/mcp.json in your workspace.

Configuration

Edit .cursor/mcp.json (create it if it doesn't exist):

{
  "mcpServers": {
    "pipelex": {
      "command": "uv",
      "args": [
        "--directory",
        "/absolute/path/to/pipelex-mcp",
        "run",
        "python",
        "-m",
        "server.main"
      ]
    }
  }
}

Important:

• Replace /absolute/path/to/pipelex-mcp with the full path to your cloned repository

• Make sure uv is in your system PATH (verify with which uv)

Example:

{
  "mcpServers": {
    "pipelex": {
      "command": "uv",
      "args": [
        "--directory",
        "../pipelex-mcp",
        "run",
        "python",
        "-m",
        "server.main"
      ]
    }
  }
}

Usage

After configuration, simply chat with Cursor:

"Build me a pipeline to extract email addresses from text and validate them."

Cursor will automatically invoke the Pipelex MCP tools to build and execute the pipeline.

Learn more about MCP in Cursor

Claude Desktop

Claude Desktop also supports MCP through its configuration file.

Configuration

1. Open Claude Desktop Settings:

   • Click Claude in the menu bar → Settings…

   • Navigate to Developer tab

   • Click Edit Config

2. Edit the configuration file:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "pipelex": {
      "command": "uv",
      "args": [
        "--directory",
        "/absolute/path/to/pipelex-mcp",
        "run",
        "python",
        "-m",
        "server.main"
      ]
    }
  }
}

Important:

• Replace /absolute/path/to/pipelex-mcp with the full path to your cloned repository

• Make sure uv is in your system PATH (verify with which uv)

Example:

{
  "mcpServers": {
    "pipelex": {
      "command": "uv",
      "args": [
        "--directory",
        "/Users/yourname/projects/pipelex-mcp",
        "run",
        "python",
        "-m",
        "server.main"
      ]
    }
  }
}

3. Restart Claude Desktop completely

For more detailed setup instructions, see the official MCP documentation.

Other Clients

MCP is an open protocol. Any client that implements MCP can use this server. Check the MCP documentation for a full list of compatible clients.

💡 Examples

Example 1: CV-Job Matching & Interview Question Generation

Use Case: Automatically analyze how well a candidate's CV matches a job offer and generate relevant interview questions.

What it does:

1. Extracts text from a CV PDF

2. Analyzes alignment between CV and job requirements

3. Generates targeted interview questions based on gaps and strengths

4. Compiles everything into a comprehensive interview preparation document

How to use it:

Simply tell your AI agent:

"Build me a pipeline that takes a CV PDF and a job offer text, then generates interview questions based on the candidate's alignment with the role."

Then provide:

Inputs:
- CV PDF: https://example.com/sample-cv.pdf
- Job Offer Text: "GenAI Engineer at Pipelex (Paris)
  
  We're growing our team of AI engineers in Paris to build our Agentic Knowledge Framework.
  
  Key requirements:
  - Strong Python skills
  - Experience with LLMs and AI frameworks
  - Ability to ship fast and iterate
  - Knowledge of software architecture
  - Passion for building developer tools
  
  ..."

The agent will:

1. Use pipe_builder to construct the pipeline definition

2. Use pipe_runner to execute it with your inputs

3. Return structured interview questions and analysis

⚠️ Known Limitations

We are actively working on improving these aspects:

Instability

It is not stable yet. The inputs are not always correctly parsed. You might encounter errors like this:

"Pipe 'name_of_pipe' already exists in the library".

This will be fixed in the coming releases.

UV Package Manager Dependency

Currently, the MCP server requires uv to be installed and configured. While you can modify the configuration to use alternative Python package managers, the default setup relies on uv. We're working on providing more flexible installation options in future releases.

Logging Configuration

The current logging system has some limitations:

• Difficulty in redirecting logs to specific files

• Limited control over log formatting and destinations

• No built-in log rotation or management

We're planning to implement a more robust logging system in future updates.

Long-Running Pipelines

For pipelines that take longer than the MCP client timeout to complete, we need to implement a session ID system. This would allow:

• Handling of timeouts gracefully

• Resuming pipeline execution

• Status tracking for long-running operations

This feature is planned for future releases to better support extended pipeline operations.

🤝 Contributing

We welcome contributions! Please check our issues page or submit a pull request.

💬 Support

• 💬 Discord Community - Get help and share your workflows

• 🐛 GitHub Issues - Bug reports and feature requests

• 📚 Pipelex Documentation - Complete guides and references

• 🌐 Pipelex Homepage - Learn more about Pipelex

📝 License

This project is licensed under the MIT license.

"Pipelex" is a trademark of Evotis S.A.S.

© 2025 Evotis S.A.S.