Skip to main content

Getting Started

Agent Toolkit is the tooling layer of the IoWarp platform, providing comprehensive capabilities for AI agents working in scientific computing environments.

v1.0.0 Release (Beta Public Release - November 11, 2025) launches with 15+ Model Context Protocol (MCP) servers, enabling AI coding assistants (Cursor, Claude Code, VS Code) to interact with HPC resources, scientific data formats, and research datasets through natural language.

Future Roadmap: v1.2.0 and beyond will expand beyond MCP servers to include additional skills, plugins, and extensions, making Agent Toolkit a complete ecosystem for AI agent tooling.

Built by: Gnosis Research Center (GRC) at Illinois Institute of Technology
Contact: grc@illinoistech.edu
Platform: IoWarp.ai
Supported by: National Science Foundation
Technology: FastMCP 2.12, Python 3.10+, MIT licensed

Quick Start​

Install uv (if needed)​

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

Run a Server​

# List all 15 available servers
uvx agent-toolkit

# Example: Run HDF5 server
uvx agent-toolkit hdf5

This launches the server via stdio transport. Your AI assistant can now use it.

Add to Your AI Assistant​

Cursor: Edit ~/.cursor/mcp.json

{
  "mcpServers": {
    "hdf5": {"command": "uvx", "args": ["agent-toolkit", "hdf5"]}
  }
}

Claude Code:

claude mcp add hdf5 -- uvx agent-toolkit hdf5

VS Code: Add to settings

"mcp": {
  "servers": {
    "hdf5": {"type": "stdio", "command": "uvx", "args": ["agent-toolkit", "hdf5"]}
  }
}

Restart your editor. The MCP server tools will be available in AI assistant context.

What Can You Do?​

With HDF5 MCP (27 tools)​

Work with HDF5 scientific data files:

  • Explore file structure and datasets
  • Read full or partial datasets
  • Access metadata and attributes
  • Parallel batch processing
  • Stream large datasets efficiently
  • AI-powered structure analysis

Example prompt: "Open simulation.h5 and show me the temperature dataset structure"

With Slurm MCP (13 tools)​

Manage HPC cluster jobs:

  • Submit jobs with resource specifications
  • Monitor job status and queue position
  • Retrieve job output
  • Allocate interactive nodes
  • Query cluster information

Example prompt: "Submit train.py to Slurm with 32 cores, 64GB RAM, 24 hours"

With Pandas MCP (15 tools)​

Process and analyze tabular data:

  • Load CSV, Excel, Parquet, HDF5
  • Statistical analysis and correlations
  • Data cleaning and transformation
  • Time series operations
  • Save to multiple formats

Example prompt: "Load data.csv, clean missing values, compute statistics by group"

With ArXiv MCP (13 tools)​

Search and retrieve research papers:

  • Search by author, title, keywords, date
  • Download PDFs
  • Export citations to BibTeX
  • Find similar papers

Example prompt: "Find recent diffusion model papers and export top 5 to BibTeX"

Other Servers​

  • ADIOS (5 tools) - Read ADIOS2 BP5 files
  • Darshan (10 tools) - Analyze I/O performance traces
  • Lmod (10 tools) - Manage environment modules
  • Plot (6 tools) - Generate plots from CSV data
  • Compression - GZIP compression
  • Jarvis (27 tools) - Data pipeline management
  • ChronoLog (4 tools) - Distributed logging
  • NDP (3 tools) - Dataset discovery via CKAN
  • Node Hardware (11 tools) - System monitoring
  • Parallel Sort (13 tools) - Large file sorting
  • Parquet - Parquet file operations

Browse all servers →

Architecture​

Each MCP server is an independent Python package with its own dependencies. The agent-toolkit launcher uses uvx to run servers in isolated environments.

Repository structure:

agent-toolkit/
├── src/agent_toolkit/       # Unified launcher (180 lines, Click only)
├── agent-toolkit-mcp-servers/     # 15 independent server packages
│   ├── hdf5/              # v1.0.0 - 27 tools, FastMCP 2.12.5, h5py 3.15.1
│   ├── pandas/            # v1.0 - 15 tools
│   ├── slurm/             # v1.0 - 13 tools
│   └── ...                # 12 more servers
└── pyproject.toml         # Launcher config only

Design benefits:

  • Dependency isolation (each server has own requirements)
  • Independent development (students work on separate servers)
  • Unified user experience (single uvx agent-toolkit <name> command)
  • Auto-discovery (launcher scans for servers via pyproject.toml)

HDF5 MCP - Reference Implementation​

HDF5 MCP v1.0.0 demonstrates MCP best practices. Study this server as a template:

Dependencies: FastMCP 2.12.5, h5py 3.15.1, numpy 2.3.4 (latest as of Oct 2025)

MCP Protocol Features:

  • 27 tools with complete annotations (title, readonly, destructive, idempotent, openworld hints)
  • Context API: progress reporting, AI-powered insights via LLM sampling
  • 3 resource URIs with templates
  • 4 workflow prompts for guided analysis

Code Quality:

  • Full type coverage (MyPy checked)
  • 10 passing tests with realistic fixtures
  • Educational demo script with sample climate data
  • Comprehensive documentation (README, TOOLS.md, ARCHITECTURE.md, EXAMPLES.md, TRANSPORTS.md)

Performance:

  • LRU caching (100-1000x speedup on repeated queries)
  • Parallel processing via ThreadPoolExecutor (4-8x speedup)
  • Streaming for datasets larger than RAM
  • Adaptive performance monitoring

Location: agent-toolkit-mcp-servers/hdf5/

Try the demo:

cd agent-toolkit-mcp-servers/hdf5/examples
uv run python create_demo_data.py
uv run python demo_script.py

View HDF5 documentation →

Development​

Clone Repository​

git clone https://github.com/iowarp/agent-toolkit.git
cd agent-toolkit

Work on a Server​

cd agent-toolkit-mcp-servers/hdf5
uv sync --all-extras --dev
uv run pytest tests/ -v
uv run hdf5-mcp

Add a New Server​

  1. Create directory: agent-toolkit-mcp-servers/my-server/
  2. Add pyproject.toml with entry point
  3. Implement server with FastMCP decorators
  4. Add tests
  5. Launcher auto-discovers it

See CONTRIBUTING.md for complete guide.

Support​

Institutional Links:

Roadmap & Vision​

v1.0.0 (Beta Public Release - November 11, 2025)

  • 15+ MCP servers for scientific computing
  • Unified launcher with auto-discovery
  • Comprehensive documentation and examples

v1.2.0+ (Future Releases)

  • Additional agent skills beyond MCP
  • Plugin system for extensibility
  • Agent extensions and integrations
  • Expanded tooling ecosystem

Agent Toolkit is evolving from a collection of MCP servers into a comprehensive platform for AI agent tooling, all within the IoWarp ecosystem.

Citation​

If you use Agent Toolkit in your research:

Agent Toolkit: Tools, Skills, Plugins, and Extensions for AI Agents
Part of the IoWarp Platform
Gnosis Research Center (GRC), Illinois Institute of Technology
https://iowarp.ai/agent-toolkit/
https://github.com/iowarp/agent-toolkit

Funding: This work is supported in part by the National Science Foundation.