scigantic-mcp 0.1.0

MCP server for the Scigantic catalog of public scientific data archives, with LLM-ready schema cards

scigantic-mcp

A Model Context Protocol server that exposes the Scigantic catalog of public scientific data archives to any MCP client.

It is built to drop into Kiro for Life Sciences alongside its domain database servers. Where each Kiro server wraps one domain's APIs (genomics, proteomics, structural, …), Scigantic is the cross-domain launchpad: 5,000+ curated public archives spanning every domain, each with an LLM-ready schema card (file format, columns, sample rows/headers, inlined READMEs/data dictionaries, and a copy-paste starter cell) so an agent can understand a dataset's structure before downloading anything.

It is discovery-only and zero-config — every tool calls public, read-only Scigantic endpoints, so there is no API key to set up.

Requires Python 3.10+ and MCP SDK 2.x.

Prefer the hosted server if your client speaks HTTP. Scigantic also runs a remote MCP server at https://api.scigantic.com/mcp (no auth, nothing to install, and it carries two extra tools): claude mcp add --transport http scigantic https://api.scigantic.com/mcp. This package exists for clients that launch stdio servers, such as Kiro.

Tools

Tool	What it does
search_archives(query, category?, limit?)	Natural-language search across the whole catalog.
get_archive(id)	Full metadata for one archive.
get_schema_card(id)	The compact schema card — the fastest way to learn a dataset's structure.
get_data_access(id, language?)	How to load the dataset in your own environment: storage location + copy-paste code snippets.
list_archive_files(id, limit?)	A sample of the files/objects in the archive.

Prompts (guided workflows)

These surface as slash commands in Claude Code (/mcp__scigantic__<name>):

Prompt	What it does
explore_dataset(topic)	Search → inspect schema cards → recommend the best dataset → offer load code.
start_analysis(archive_id, goal?)	Pull schema card + data-access snippet for an archive and outline an analysis plan.

Install & register in Kiro

Add an entry under mcpServers in ~/.kiro/settings/mcp.json.

Option A — uvx (zero-install, recommended):

{
  "mcpServers": {
    "scigantic": {
      "command": "uvx",
      "args": ["scigantic-mcp"]
    }
  }
}

Option B — install into a venv (matches the Kiro servers' own mcp.json form):

python3 -m venv .venv && . .venv/bin/activate
pip install scigantic-mcp                 # or: pip install /path/to/scigantic-mcp

{
  "mcpServers": {
    "scigantic": {
      "command": "/path/to/.venv/bin/scigantic-mcp",
      "env": {
        "SCIGANTIC_API_URL": "https://api.scigantic.com"
      }
    }
  }
}

Works the same in Claude Desktop / Claude Code (claude mcp add scigantic -- uvx scigantic-mcp) or any MCP client that launches stdio servers.

Configuration

Env var	Default	Purpose
SCIGANTIC_API_URL	https://api.scigantic.com	API base (set to https://staging-api.scigantic.com for staging).
SCIGANTIC_API_ORIGIN	https://scigantic.com	Origin header → selects the public (default) catalog tenant.

Develop & test

The tool/client layer has no mcp dependency, so those tests run on any Python with httpx and need no network (mocked transport):

python3 tests/test_tools.py     # or: pytest

tests/test_server_import.py covers the wiring layer — that the server module imports, and that the registered tools and prompts are the expected set. It needs the mcp SDK installed (Python ≥ 3.10) and is skipped otherwise:

pip install -e '.[test]' && pytest

Keep it that way: the tool tests skip server.py on purpose, so an SDK breaking change is invisible to them. All 12 passed while the server could not import at all under SDK 2.x, which is what test_server_import.py now guards against.

Roadmap

• Richer discovery for agents — structured tool outputs and MCP resources (attach an archive + its schema card as durable context).
• Upstream inclusion as life-sciences-scigantic in aws-samples/sample-kiro-power-life-sciences (see docs/kiro-upstream-pr.md).
• Hosted compute is intentionally not exposed here. Scigantic's notebooks are interactive (a JupyterLab URL a human opens); handing an external agent that URL is a dead end. The agent-to-agent path is get_data_access — the caller runs the analysis in its own environment. Letting Scigantic execute code for an agent (run against an ephemeral kernel with the dataset mounted, return outputs) is a separate capability the platform would need to build first.

License

MIT-0.