Science data pack

A skill for reading scientific data without changing it, and two clio-kit MCP servers for the analysis Clio's built-in data tool does not do. The skill answers with data first: structure, CF attributes, hyperslabs and summary statistics for HDF5, NetCDF and Parquet, plus CSV and JSON, read-only and with no Python. It moves to parquet (row slices, filtered column aggregates) and pandas (group-by, correlations, hypothesis tests, pivots) only for questions data cannot answer. A third server, hdf5, is an optional operator entry (below), and h5py through run_script is the last resort. Installing the pack launches nothing.

clio-coder library install plugin:science-data --project
clio-coder mcp list                 # shows each server and its declared per-tool classes
clio-coder mcp trust parquet
clio-coder mcp trust pandas

Read-only by default

mcp.json gives every tool an action class (toolActionClasses). clio-coder mcp list and clio-coder mcp trust print the map before you trust a server, and the trust record covers it, so changing a class in the installed pack makes the server untrusted again.

Server	Asks first (`destructive`)	Everything else
`hdf5` (operator entry below)	`open_file` (its `mode` can be `w`, `r+` or `a`), `export_dataset` (writes a file)	`read`
`pandas`	`save_data`, and every tool that writes a CSV beside its input: `clean_data`, `filter_data`, `groupby_operations`, `handle_missing_data`, `merge_datasets`, `optimize_memory`, `pivot_table`, `time_series_operations`	`read`
`parquet`	none	`read`

A read tool runs without asking at default and yolo. A destructive tool asks at default. The safety judge reviews the ask first, reading only your prompt, so a headless run that asks for a grouped file can still write one; a headless run denies what the judge does not allow. Trust a server without --action-class: the per-tool map already classifies every tool it ships, and a tool a later clio-kit release adds stays unknown, which asks.

The pack needs Clio Coder 0.7.0 or later, the first release whose plugin mcp.json accepts toolActionClasses.

HDF5: an operator entry, not a pack declaration

The clio-kit hdf5 server writes its index files to data/ under its working directory on every launch, and that path is fixed. A plugin's server runs in the installed plugin tree, so those writes would change the pinned package and stop it loading. Declare hdf5 yourself in the user file <Clio config directory>/mcp.yaml instead, with an absolute cwd that exists and that you do not mind it writing to, such as a directory under Clio's state directory (clio-coder paths prints it). The classes match the table above:

version: 1
servers:
  - id: hdf5
    command: clio-kit
    args: [mcp-server, hdf5]
    cwd: /absolute/path/to/clio-state/mcp/hdf5
    timeoutMs: 120000
    toolActionClasses:
      analyze_dataset_structure: read
      close_file: read
      export_dataset: destructive
      find_similar_datasets: read
      get_by_path: read
      get_chunks: read
      get_dtype: read
      get_filename: read
      get_mode: read
      get_shape: read
      get_size: read
      hdf5_aggregate_stats: read
      hdf5_batch_read: read
      hdf5_parallel_scan: read
      hdf5_stream_data: read
      identify_io_bottlenecks: read
      list_attributes: read
      list_available_hdf5_files: read
      list_keys: read
      open_file: destructive
      optimize_access_pattern: read
      read_attribute: read
      read_full_dataset: read
      read_partial_dataset: read
      refresh_hdf5_resources: read
      suggest_next_exploration: read
      visit: read

A user entry is trusted by authorship, so no mcp trust step is needed.

Requirements and environment

clio-kit must be on PATH (uv tool install clio-kit). Clio passes an MCP server only PATH, HOME, USER, LOGNAME, LANG, LC_ALL, the temp and XDG_* directories and a few platform variables. The servers need nothing else. CLIO_KIT_CACHE_DIR from your shell does not reach them, so their cache is $XDG_CACHE_HOME/clio-kit or ~/.cache/clio-kit.

Offline behavior

The servers read local files and make no network calls while they run. The data tool's HDF5 and Parquet readers download once from the npm registry on first use and are cached; offline with a cold cache it refuses and names the cache path. clio-kit runs each server from its own lock file in a uv environment. The first launch downloads that environment from PyPI and fails without network. Every later launch from the warm cache works offline. A plugin's servers run with the installed plugin directory as their working directory, so the skill passes absolute file paths. On an air-gapped node, warm the cache once on a connected machine that shares the home directory, or copy the cache over.

Licenses and terms

The pack is Apache-2.0 and bundles no code or data. It queries no external data service, so no data terms of use apply. clio-kit is BSD-3-Clause; its servers pull h5py, NumPy, pandas and PyArrow under their own licenses into the uv environment.