ICRISAT Data Hub MCP Server

# ICRISAT Data Hub

A lightweight, federation-first data hub for ICRISAT teams, modelled on the CGIAR Climate Data Hub asset-mapping approach:

teams drop files  ──►  inbox (watched)  ──►  ingest pipeline  ──►  catalogue  ──►  MCP server  ──►  any LLM
   (no skills needed)                      (auto: read, profile,               (SQLite + assets.json)
                                           register, flag gaps)

Phase 1 supports CSV, Excel (.xlsx/.xls), and Word (.docx) files. External data is registered as YAML pointers (federation — the hub points to data where it lives rather than copying it).

For teams: how to share data (30 seconds, no skills needed)

Option A — drop folder. Copy your file into data/inbox/ (this can be a shared network drive or a OneDrive-synced folder). That's it. The pipeline picks it up automatically, catalogues it, and flags anything it couldn't figure out for review.

Option B — web page. Run scripts\start_web.cmd once, open http://localhost:8010, drag & drop, optionally fill 5 small fields.

Option C — with metadata. Drop a tiny sidecar next to your file:

yield_trials_2025.csv
yield_trials_2025.meta.yaml     # title, team, owner, description, tags, ...

or one metadata.yaml for a batch of files. Incomplete uploads get a pre-filled *_metadata_review.yaml written next to the processed file — complete it and drop it back in the inbox (or edit via the review CLI).

Related MCP server: mix_server

Setup

python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt

Running the pieces

All python -m hub... commands run from the project root with PYTHONPATH=src (the scripts set this for you).

Connecting an LLM (MCP)

Claude Desktop / opencode (stdio) — add to your MCP config:

{
  "mcpServers": {
    "icrisat-data-hub": {
      "command": "C:\\Users\\HDesalegn\\ICRISAT\\icrisat-data-hub\\.venv\\Scripts\\python.exe",
      "args": ["-m", "hub.mcp_server.server"],
      "env": { "PYTHONPATH": "C:\\Users\\HDesalegn\\ICRISAT\\icrisat-data-hub\\src" }
    }
  }
}

Other machines/LLMs (HTTP) — start with --http and point clients at http://<host>:8100/mcp.

MCP tools

YAML source pointers (federation)

Each external dataset/API/database gets one small YAML in sources/ — see sources/README.md and sources/examples/. The hub registers and exposes these through the MCP server without copying the data.

GitHub Pages catalogue app

A static dashboard lives in docs/ and is served at https://henokd11.github.io/icrisat-data-hub/. It reads docs/data/*.json — regenerate and commit them after the catalogue changes:

set PYTHONPATH=src
.venv\Scripts\python -m hub.pages
git add docs/data && git commit -m "Refresh catalogue app data" && git push

Layout

config/hub.yaml          hub-wide settings (paths, vocab, ports)
data/inbox/              drop zone (watched)
data/processed/          ingested files, organised by team/month
data/failed/             rejected files + .error.txt
data/catalog/            catalog.db (SQLite) + assets.json (exported catalogue)
sources/                 YAML pointers to federated external data
src/hub/                 pipeline, watcher, web app, MCP server
tests/                   pytest suite