CLIO Kit

CLIO Kit - Part of the IoWarp platform's tooling layer for AI agents. A comprehensive collection of tools, skills, plugins, and extensions. It ships 22 Model Context Protocol (MCP) servers for scientific computing and enables AI agents to interact with HPC resources, scientific data formats, and research datasets.

Website | IOWarp

Chat with us on Zulip or join us

Developed by <img src="https://grc.iit.edu/img/logo.png" alt="GRC Logo" width="18" height="18"> Gnosis Research Center

❌ Without CLIO Kit

Working with scientific data and HPC resources requires manual scripting and tool-specific knowledge:

• ❌ Write custom scripts for every HDF5/Parquet file exploration
• ❌ Manually craft Slurm job submission scripts
• ❌ Switch between multiple tools for data analysis
• ❌ No AI assistance for scientific workflows
• ❌ Repetitive coding for common research tasks

✅ With CLIO Kit

AI agents handle scientific computing tasks through natural language:

• ✅ "Analyze the temperature dataset in this HDF5 file" - HDF5 MCP does it
• ✅ "Submit this simulation to Slurm with 32 cores" - Slurm MCP handles it
• ✅ "Find papers on neural networks from ArXiv" - ArXiv MCP searches
• ✅ "Plot the results from this CSV file" - Plot MCP visualizes
• ✅ "Optimize memory usage for this pandas DataFrame" - Pandas MCP optimizes
• ✅ "Find all documents where pressure exceeds 200 kPa" - Agentic Search retrieves

One unified interface. 22 MCP servers. Hybrid search engine. 150+ specialized tools. Built for research.

CLIO Kit is part of the IoWarp platform's comprehensive tooling ecosystem for AI agents. It brings AI assistance to your scientific computing workflow—whether you're analyzing terabytes of HDF5 data, managing Slurm jobs across clusters, or exploring research papers. Built by researchers, for researchers, at Illinois Institute of Technology with NSF support.

Part of IoWarp Platform: CLIO Kit is the tooling layer of the IoWarp platform, providing skills, plugins, and extensions for AI agents working in scientific computing environments.

One simple command. Production-ready, fully typed, BSD-3-Clause licensed, and live-tested in real HPC environments.

🚀 Quick Installation

One Command for Any Server

# Install the released CLI into its own persistent tool environment
uv tool install 'clio-kit==2.4.3'
# If uv reports that its executable directory is not on PATH:
uv tool update-shell

# List all 23 available MCP servers
clio-kit mcp-servers

# Run any installed server
clio-kit mcp-server hdf5
clio-kit mcp-server pandas
clio-kit mcp-server slurm

# Agentic search — hybrid retrieval for scientific corpora
clio-kit search serve               # Start search API server
clio-kit search query --namespace local_fs --q "pressure > 200 kPa"

# AI prompts also available
clio-kit prompts                    # List all prompts
clio-kit prompt code-coverage-prompt # Use a prompt

uv tool install keeps CLIO Kit in a persistent, isolated tool environment. Use uvx --from 'clio-kit==2.4.3' clio-kit ... only for a temporary, one-shot invocation.

Released clio-kit wheels execute each embedded MCP server from that server's shipped uv.lock. The launcher uses a source-and-lock-addressed environment under the user cache, installs only production dependencies, and refuses to resolve an embedded server whose lock is missing. The --branch launcher option is an explicit development path and is not an immutable release-artifact path.

The root wheel also ships machine-readable user contracts for the locked JARVIS, SLURM, Spack, and Scientific Catalog servers. These artifacts are generated from real stdio tools/list exchanges and include canonical SHA-256 digests for downstream federation gates:

clio-kit mcp-contracts
clio-kit mcp-contract clio-kit-jarvis-user-v3.5
clio-kit mcp-contract clio-kit-slurm-user-v3
clio-kit mcp-contract clio-kit-spack-user-v2.1
clio-kit mcp-contract clio-kit-scientific-catalog-user-v1.1

Add to your Cursor ~/.cursor/mcp.json:

{
  "mcpServers": {
    "hdf5-mcp": {
      "command": "clio-kit",
      "args": ["mcp-server", "hdf5"]
    },
    "pandas-mcp": {
      "command": "clio-kit",
      "args": ["mcp-server", "pandas"]
    },
    "slurm-mcp": {
      "command": "clio-kit",
      "args": ["mcp-server", "slurm"]
    }
  }
}

See Cursor MCP docs for more info.

# Add HDF5 MCP
claude mcp add hdf5-mcp -- clio-kit mcp-server hdf5

# Add Pandas MCP
claude mcp add pandas-mcp -- clio-kit mcp-server pandas

# Add Slurm MCP
claude mcp add slurm-mcp -- clio-kit mcp-server slurm

See Claude Code MCP docs for more info.

Add to your VS Code MCP config:

"mcp": {
  "servers": {
    "hdf5-mcp": {
      "type": "stdio",
      "command": "clio-kit",
      "args": ["mcp-server", "hdf5"]
    },
    "pandas-mcp": {
      "type": "stdio",
      "command": "clio-kit",
      "args": ["mcp-server", "pandas"]
    }
  }
}

See VS Code MCP docs for more info.

Edit claude_desktop_config.json:

{
  "mcpServers": {
    "hdf5-mcp": {
      "command": "clio-kit",
      "args": ["mcp-server", "hdf5"]
    },
    "arxiv-mcp": {
      "command": "clio-kit",
      "args": ["mcp-server", "arxiv"]
    }
  }
}

See Claude Desktop MCP docs for more info.

Available Packages

The version below is each MCP server's agent-facing contract version, not the containing clio-kit wheel version. JARVIS 3.4 and SLURM 3.0 have contracts redesigned for agent use. Spack is at 2.1, while the other contracts retain their existing 2.x identities until a focused upgrade.

The Spack install contract makes concretization explicit: reuse=true passes spack install --reuse, while reuse=false passes spack install --fresh. Agents should discover first, install only when needed, then pass the exact spack_locate result to JARVIS for runtime loading. A find with no installed match is normal typed data (count=0, packages=[]); locate reports the distinct not_installed semantic, while real Spack failures remain errors.

Agentic Search

Hybrid retrieval engine for scientific corpora — combines lexical (BM25), vector, graph, and scientific search (numeric range, unit matching, formula targeting) over namespaced document collections. DuckDB storage, FastAPI, async job queue, OpenTelemetry tracing, Prometheus metrics.

# Start the search API server
clio-kit search serve

# Index documents from a namespace
clio-kit search index --namespace local_fs

# Query with scientific operators
clio-kit search query --namespace local_fs --q "pressure between 190 and 360 kPa"

# List indexed documents
clio-kit search list --namespace local_fs

API endpoints: /query, /jobs/index, /documents, /health, /metrics — full docs

📖 Usage Examples

HDF5: Scientific Data Analysis

"What datasets are in climate_simulation.h5? Show me the temperature field structure and read the first 100 timesteps."

Tools used: open_file, analyze_dataset_structure, read_partial_dataset, list_attributes

Slurm: HPC Job Management

"Submit simulation.py to Slurm with 32 cores, 64GB memory, 24-hour runtime. Monitor progress and retrieve output when complete."

Tools used: submit_slurm_job, check_job_status, get_job_output

ArXiv: Research Discovery

"Find the latest papers on diffusion models from ArXiv, get details on the top 3, and export citations to BibTeX."

Tools used: search_arxiv, get_paper_details, export_to_bibtex, download_paper_pdf

Pandas: Data Processing

"Load sales_data.csv, clean missing values, compute statistics by region, and save as Parquet with compression."

Tools used: load_data, handle_missing_data, groupby_operations, save_data

Plot: Data Visualization

"Create a line plot showing temperature trends over time from weather.csv with proper axis labels."

Tools used: line_plot, data_info

Agentic Search: Scientific Retrieval

"Find all chunks mentioning pressure above 200 kPa in the local_fs namespace."

CLI: clio-kit search query --namespace local_fs --q "pressure > 200 kPa"

🚨 Troubleshooting

If clio-kit mcp-server <server-name> fails:

# Verify server name is correct
clio-kit mcp-servers

# Common names: hdf5, pandas, slurm, arxiv (not hdf5-mcp, pandas-mcp)

For development or local testing:

cd clio-kit-mcp-servers/hdf5
uv sync --all-extras --dev
uv run hdf5-mcp

Install uv package manager:

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

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

# Or via pip
pip install uv

Then install CLIO Kit persistently and expose uv's tool directory:

uv tool install 'clio-kit==2.4.3'
uv tool update-shell

Team

• Gnosis Research Center (GRC) - Illinois Institute of Technology | Lead
• HDF Group - Data format and library developers | Industry Partner
• University of Utah - Research collaboration | Domain Science Partner

Sponsored By

<img src="https://www.nsf.gov/themes/custom/nsf_theme/components/molecules/logo/logo-desktop.png" alt="NSF Logo" width="24" height="24"> NSF (National Science Foundation) - Supporting scientific computing research and AI integration initiatives

we welcome more sponsorships. please contact the Principal Investigator

Ways to Contribute

• Submit Issues: Report bugs or request features via GitHub Issues
• Develop New MCPs: Add servers for your research tools (CONTRIBUTING.md)
• Improve Documentation: Help make guides clearer
• Share Use Cases: Tell us how you're using CLIO Kit in your research

Full Guide: CONTRIBUTING.md

Community & Support

• Chat: Zulip Community
• Join: Invitation Link
• Issues: GitHub Issues
• Discussions: GitHub Discussions
• Website: https://docs.iowarp.ai/
• Project: IOWarp Project