01
Why science needs its own servers
General-purpose tools do not understand hierarchical scientific formats, job schedulers, I/O traces, or domain catalogs. A server that knows how to list HDF5 groups or parse a Darshan log gives an agent precise operations instead of guesswork over raw bytes.
02
HDF5 with agents
HDF5 is a standard format for simulation and experimental data across the sciences. CLIO Kit’s HDF5 server lets an agent list groups, inspect datasets and attributes, and read slices without pulling whole files, and CLIO Core’s HDF5 VOL connector lets existing HDF5 applications place data through the runtime without code changes.
The HDF Group works on agentic HDF5 and agent skills with the IOWarp team, and at SC26 a birds-of-a-feather session covers HDF5 in AI-driven science.
03
What CLIO Kit provides
CLIO Kit is an open-source meta-marketplace of scientific MCP servers, skills, agents, hooks, and workflow plugins. Its servers cover data formats (HDF5, Parquet, ADIOS, Pandas), HPC operations (Slurm, Spack, Lmod, Jarvis, Darshan, node hardware), visualization (ParaView, Plot), geospatial and seismic data, and discovery (ArXiv, NDP, scientific catalogs, web search).
The launcher installs from PyPI as clio-kit and downloads selected components on demand. Project installers support Codex, Claude Code, OpenCode, Cursor, Antigravity, and VS Code; agents and hooks have client-specific adapters and limits.
04
Design choices
Each server is separately versioned and testable, so a facility can adopt only the servers it trusts. Component downloads are verified, and server dependencies stay isolated and locked. Tools return structured results, which makes an agent’s steps easier to check afterward.