IOWarp MCPs

Overview

This project implements a Scientific Model Context Protocol (MCP) Server using FastAPI and JSON-RPC 2.0. The server simulates various scientific computing capabilities and allows AI agents and Large Language Models (LLMs) to interact with tools and data sources in a standardized way.

Implemented MCP Capabilities

The following capabilities have been implemented:

Capability	Type	Description
`Adios`	Data	Reads data from different file types using the ADIOS2 engine.
`Arxiv`	Data	Fetches recent research papers from the Arxiv API.
`ChronoLog`	External System	Provides tools to log and retrieve data from a ChronoLog server.
`Compression`	Tool	Simulates file compression using Python's `gzip` module.
`Darshan`	Analysis	Analyzes I/O profiler trace files for performance insights.
`HDF5`	Data	Lists `.hdf5` files from a specified local directory.
`Jarvis`	Tool	Manages the full lifecycle of data-centric pipelines.
`Lmod`	Tool	Manages environment modules using the Lmod system.
`Node_Hardware`	Tool	Reports the number of CPU cores on the current system.
`Pandas`	Data	Loads and filters data from a CSV file using the `pandas` library.
`Parallel_Sort`	Tool	Simulates sorting a large text file and returns the sorted result.
`Parquet`	Data	Reads a specific column from a Parquet file using `pyarrow`.
`Plot`	Tool	Generates a plot from a local CSV file using `pandas` and `matplotlib`.
`Slurm`	Tool	Simulates Slurm-like job submission and returns a fake job ID.

Prerequisites

Python 3.10 or higher (https://www.python.org/)
uv package manager
Linux/macOS environment (for optimal compatibility)

Quick Start (Recommended)

The easiest way to use any MCP server is with the unified launcher:

# Run any server directly with uvx (no installation required)
uvx iowarp-mcps adios
uvx iowarp-mcps hdf5
uvx iowarp-mcps slurm

# List all available servers
uvx iowarp-mcps

# Run with additional arguments
uvx iowarp-mcps pandas --help

This approach automatically manages dependencies for each server in isolation.

Installation

The Scientific Mcps supports multiple installation methods:

Global Installation of all mcps together -

Clone the repository:

git clone https://github.com/iowarp/scientific-mcps.git
cd scientific-mcps

Create and activate environment:

# On Windows
python -m venv mcp-server
mcp-server\Scripts\activate 

#On macOS/Linux
python3 -m venv mcp-server
source mcp-server/bin/activate  #On macOS/Linux

Install uv:
```
pip install uv
```

You can install all MCPs at once or select them individually.

To install all MCPs:

# This installs all dependencies listed in the pyproject.toml
uv pip install --requirement pyproject.toml

To install individual or multiple MCPs:

MCP	Installation Code (`uv pip install ...`)	Documentation
`Adios`	`"git+https://github.com/iowarp/scientific-mcps.git@main#subdirectory=Adios"`	docs
`Arxiv`	`"git+https://github.com/iowarp/scientific-mcps.git@main#subdirectory=Arxiv"`	docs
`ChronoLog`	`"git+https://github.com/iowarp/scientific-mcps.git@main#subdirectory=Chronolog"`	docs
`Compression`	`"git+https://github.com/iowarp/scientific-mcps.git@main#subdirectory=Compression"`	docs
`Darshan`	`"git+https://github.com/iowarp/scientific-mcps.git@main#subdirectory=Darshan"`	docs
`HDF5`	`"git+https://github.com/iowarp/scientific-mcps.git@main#subdirectory=HDF5"`	docs
`Jarvis`	`"git+https://github.com/iowarp/scientific-mcps.git@main#subdirectory=Jarvis"`	docs
`Lmod`	`"git+https://github.com/iowarp/scientific-mcps.git@main#subdirectory=lmod"`	docs
`Node_Hardware`	`"git+https://github.com/iowarp/scientific-mcps.git@main#subdirectory=Node_Hardware"`	docs
`Pandas`	`"git+https://github.com/iowarp/scientific-mcps.git@main#subdirectory=Pandas"`	docs
`Parallel_Sort`	`"git+https://github.com/iowarp/scientific-mcps.git@main#subdirectory=Parallel_Sort"`	docs
`Parquet`	`"git+https://github.com/iowarp/scientific-mcps.git@main#subdirectory=parquet"`	docs
`Plot`	`"git+https://github.com/iowarp/scientific-mcps.git@main#subdirectory=Plot"`	docs
`Slurm`	`"git+https://github.com/iowarp/scientific-mcps.git@main#subdirectory=Slurm"`	docs

Tip: You can install multiple MCPs in a single command by listing them one after another (e.g., uv pip install "adios-mcp..." "arxiv-mcp...").

Running the Server with different types of Clients:

Running the Universal Client (`wrp_chat`)

This repository includes a universal client, bin/wrp.py, that allows you to interact with any MCP server using natural language. It supports multiple LLM providers (Gemini, OpenAI, Claude, Ollama).

For a quick Gemini setup -

Install Client Dependencies:

# From the root directory
uv pip install -r bin/requirements.txt

Configure API Keys: Your API keys for providers like Gemini, OpenAI, or Anthropic are managed in the configuration files.

For long-term use, open the relevant pre-configured file in bin/confs (e.g., Gemini.yaml) and enter your key directly:
```
# In bin/confs/Gemini.yaml
LLM:
Provider: Gemini
api_key: your-gemini-api-key # <-- ADD KEY HERE
model_name: gemini-1.5-flash
```
For one-time use, you can use environment variables. First, export the key in your terminal:
```
# On macOS/Linux
export GEMINI_API_KEY="your-gemini-api-key"
# On Windows
$env:GEMINI_API_KEY="your-gemini-api-key"
```
Run the Client: To run the client, execute the wrp script from your terminal, specifying a configuration file with the --conf flag.

Example for Gemini:
```
python bin/wrp.py --conf=bin/confs/Gemini.yaml
```
For Additional Troubleshooting & Debugging use verbose:

python bin/wrp.py --conf=bin/confs/Gemini.yaml --verbose

Running a specific Mcp directly. Run the Mcp Server directly:
```
cd Adios     # Jarvis or any other specific mcp.
uv run adios-mcp     # change the mcp name ex. jarvis-mcp   
```
This will create a .venv/ folder, install all required packages, and run the server directly.

Running the Server on Claude Command Line Interface Tool.

Install the Claude Code using NPM, Install NodeJS 18+, then run:

npm install -g @anthropic-ai/claude-code

Running the server:

claude add mcp jarvis -- uv --directory ~/scientific-mcps/Jarvis run jarvis-mcp

Running the Server on open source LLM client (Claude, Copilot, etc.)

To add the Adios MCP Put the following in settings.json of any open source LLMs like Claude or Microsoft Co-pilot:

"adios-mcp": {
    "command": "uv",
    "args": [
        "--directory",
        "path/to/directory/src/adiosmcp/",
        "run",
        "server.py"
    ]
}

Project Structure

iowarp-mcps/
├── mcps/                    # Auto-discovered MCP servers
│   ├── Adios/
│   ├── Arxiv/
│   ├── Chronolog/
│   ├── Compression/
│   ├── HDF5/
│   ├── Jarvis/
│   ├── Node_Hardware/
│   ├── Pandas/
│   ├── Parallel_Sort/
│   ├── parquet/
│   ├── Plot/
│   ├── Slurm/
│   └── lmod/
├── src/
│   └── iowarp_mcps/         # Unified launcher
│       └── __init__.py
├── bin/
│   ├── wrp.py
│   ├── README.md
│   ├── instructions.md
│   └── ...
└── ...

Adding New MCPs

To add a new MCP server:

Create directory: Add your server to mcps/YourNewServer/
Add pyproject.toml: Include entry point like your-server-mcp = "module:main"
That's it! The launcher will auto-discover it

No manual mapping required - the system automatically finds all servers in the mcps/ folder.

Usage

Unified Launcher (Recommended)

Use the unified launcher for the simplest experience:

# Run any server directly
uvx iowarp-mcps <server-name>

# Examples:
uvx iowarp-mcps adios
uvx iowarp-mcps hdf5
uvx iowarp-mcps slurm

# List available servers
uvx iowarp-mcps

Individual Server Usage

To run any MCP server directly or learn more about its specific capabilities, navigate into its directory and follow the instructions in its local README.md.

Overview​

Implemented MCP Capabilities​

Prerequisites​

Quick Start (Recommended)​

Installation​

Running the Server with different types of Clients:​

Running the Universal Client (wrp_chat)​

Running the Server on Claude Command Line Interface Tool.​

Running the Server on open source LLM client (Claude, Copilot, etc.)​

Project Structure​

Adding New MCPs​

Usage​

Unified Launcher (Recommended)​

Individual Server Usage​