Skip to main content
All tutorials

Scientific data 3 min read

Explore an HDF5 file

Install the scientific I/O workflow, invoke its dataset skill, and watch your agent inspect a file through the HDF5 MCP.

Written for
CLIO Kit 2.11
Works in
Claude Code · Codex · Clio Coder · OpenCode
Basis
Recorded sessions · 7 Oct 2026
Clio Coder loads dataset-explore and calls the clio-hdf5 tools through its gateway.
Clio Coder loads dataset-explore and calls the clio-hdf5 tools through its gateway.Clio Coder · 0.6.1

Install a workflow, open your agent, and ask what is inside a scientific data file. This walkthrough uses the clio-scientific-io components, the dataset-explore skill and the HDF5 MCP. You finish with a short report you can check against the input.

The screenshots come from real sessions recorded on 7 October 2026 in disposable folders: Claude Code 2.1.293 (Opus 5.5), Codex 0.160.1 (GPT-6-Astra), Clio Coder 0.6.1 (Claude Opus 5.5) and OpenCode 1.18.35 (GPT-5.6 Terra). Each client ran in its permission-bypass mode inside those folders; keep approvals on in your own projects. Pick your client once and every tutorial follows it.

1. Create a file you can check​

First install the CLIO Kit launcher from your checkout. Make a new folder, download create_dataset.py into it, and run:

mkdir hdf5-demo
cd hdf5-demo
uv run --no-project --with h5py python create_dataset.py

The script creates simulation.h5 and refuses to replace an existing file. It holds a five-element time axis in seconds and a (5, 4, 3) temperature array in kelvin. The first three temperatures are 273.15, 274.15 and 275.15, so you can check the agent's answer.

2. Install, ask and follow the tool calls​

Replace /path/to/clio-kit with your checkout.

clio-kit plugin install clio-scientific-io \
  --root /path/to/clio-kit --client claude-code --project "$PWD"
claude

The installer adds three skills to .claude/skills and four MCP servers to .mcp.json: HDF5, ADIOS, Parquet and Compression. Trust the folder, enable the project MCP servers when asked, then paste:

Use /dataset-explore and read its installed instructions. Inspect simulation.h5 read-only using the clio-hdf5 MCP server. Confirm the open file, list datasets, shapes, dtypes and units. Read only the first three temperature values after checking the shape. Close the file. Save a concise dataset-notes.md and show the findings. Use MCP calls for inspection, not a substitute Python reader. Work only in this demo folder; no network or delegation.

Claude loads the skill, then calls the HDF5 tools one by one. Press Ctrl+O for the detailed transcript shown here:

Claude Code transcript: the dataset-explore skill loads, then clio-hdf5 MCP calls open the file, check shapes, dtypes and attributes, read a 1×1×3 slice and close the file

Claude Code reports both datasets with shapes, dtypes and units, and the three temperatures 273.15, 274.15 and 275.15 K

The skill's useful part is the order: check structure and units before reading values. Every client asked for a start="0,0,0", count="1,1,3" slice instead of pulling the whole array into the conversation.

Select a screenshot to open it at full size. These are captures of the actual terminal sessions, not recreated conversations.

3. Check the result​

cat dataset-notes.md

Your report should identify:

DatasetShapeTypeUnits
/time(5,)float64seconds (s)
/temperature(5, 4, 3)float64kelvin (K)

Check that the sample is 273.15, 274.15, 275.15 and that the file was closed. After the four recorded sessions we checked each saved report against these known values. A correct-looking report alone is not that check.

For your own file, keep the same request but use its path. If it is large, inspect shape and dtype first and use the large-data-read skill for bounded reductions.

If the MCP is unavailable, run clio-kit doctor --server hdf5 --connect, check the client's PATH and reload the project. An installed skill does not replace a missing MCP connection.

Next: archive and summarize a dataset, summarize and plot a CSV, or contribute your own plugin.