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.

Since 0.3.0 it also carries six surechembl_* tools for patent chemistry, backed by SureChEMBL (EMBL-EBI, 31M compounds extracted from 45M patents) through the scigantic-surechembl library: look a compound up by any identifier, list the patents that mention it, read a patent, list the structures it discloses, search patents with Solr syntax, and run structure searches. Also public, read-only, no key.

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

Patent chemistry (SureChEMBL)

Identifiers are preserved verbatim in the output (SureChEMBL's attribution terms ask that SCHEMBL ids and publication numbers be kept), and each result links to surechembl.org. SureChEMBL data is CC BY 4.0.

Prompts (guided workflows)

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

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

Stability

This package is a thin client over the public Scigantic REST API. That API is not versioned and may change without notice — if a response shape moves, a pinned older release of this package can break. Pin a version you have tested, and open an issue if a tool starts returning something unexpected.

The MCP tool names and their arguments are treated as the stable surface, and will not change without a minor version bump.

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.
• 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.