CLIO Arxiv

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 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

Install in Cursor

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.

Install in Claude Code

# 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.

Install in VS Code

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.

Install in Claude Desktop

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.

| `spack` | 2.1.0 | Package Management | Structured package discovery, installation, and location | `clio-kit mcp-s