data-aggregator-mcp

data-aggregator-mcp is an MCP server that lets AI agents search 21 public research-data sources at once and get back one deduplicated list.

Install

Claude Code:

claude mcp add data-aggregator -- uvx data-aggregator-mcp

One-click install for VS Code or Cursor:

Any other MCP client: run uvx data-aggregator-mcp as a stdio server.

{
  "mcpServers": {
    "data-aggregator": {
      "command": "uvx",
      "args": ["data-aggregator-mcp"],
      "env": { "NCBI_API_KEY": "your-optional-key" }
    }
  }
}

pip install data-aggregator-mcp
data-aggregator-mcp        # or: python -m data_aggregator_mcp

What you can ask

"Find transcriptomes for Orobanche aegyptiaca." The species was renamed; NCBI Taxonomy maps it to Phelipanche aegyptiaca, and the sources that accept synonyms match both names (the demo above).

"What data came out of this paper?"

resolve("pubmed:40098680")
  links  geo:GSE284240, bioproject:PRJNA1198054, and 18 sra: experiments
  files  PMC11910882.xml   (open-access full text)

"Can I train a model on this dataset?"

resolve("hf:scikit-learn/iris", use="ml-training")
  license_compat  ALLOW: "CC0-1.0 grants the permission(s) required for
                  ml-training: commercial-use, modifications"

The verdict is read from the record's license metadata; it is advice, not a legal opinion. resolve also renders citations in any CSL style, checks Crossref for retractions, scores FAIRness, and exports Croissant or RO-Crate. relate reports how a set of records connect (a shared accession, DOI or explicit link).

How a search runs

A source that fails is named in the result's errors, never dropped silently.

What each source can fetch, and which checksum it is verified against: docs/sources.md.

What the benchmark shows

An agent using this server was compared with the same agent using web search, on held-out dataset-finding tasks scored by recall against independently built answer keys, 3 runs per task. Mean recall, server against web search: Claude Haiku 5.5 0.47 against 0.32 (higher on 8 of 10 tasks), Sonnet 5.5 0.54 against 0.52 (5 of 10), Opus 5.5 0.72 against 0.58 (6 of 10). With Haiku the server met round 5's pre-registered "better" rule (higher on at least 8 of 10 tasks, interval above 0) but not the stricter bar for a win across model tiers (9 of 10, Holm-adjusted sign test, p = 0.22); Sonnet and Opus met neither. With Opus it returned fewer wrong entries and answered in about a third of the time; with Sonnet it was faster on 9 of 10 tasks; with Haiku it cost less on 9 of 10.

The two latest Opus 5.5 rounds (2026-10-19, rules written before the runs):

In round 5 this server was higher on 6 of 10 tasks, where the pre-set rule needed 8. Web search hit the 15-minute limit in 11 of 30 runs, and a run that times out scores 0. In runs that finished, web search found more (0.90). The server answered in 203 s against 599 s (median), with fewer tool calls and fewer wrong entries (precision 0.95 against 0.62). It cost more per run than web search ($2.08 against $1.51, API-equivalent).

We chose the tasks and ran the benchmark ourselves; an outside re-run is welcome. Method and controls: benchmarks/h2h; the Haiku and Sonnet runs: 2026-10-20 re-run.

Compared with other servers

— means the project's README doesn't describe it (READMEs read 2026-10-04). Where they are ahead: BioMCP reaches clinical trials, ChEMBL and AlphaFold; ToolUniverse covers far more ground overall; paper-search-mcp covers far more literature sources; Mobus also compares datasets and checks schema compatibility. More: docs/POSITIONING.md.

Reference and development

Every tool and parameter, the HTTP transport (--transport http) and the environment variables: docs/reference.md. Sources that need every word of a query, and how they are searched: every-word sources. Tests, linting and the live API probes: CONTRIBUTING.md.

License

MIT. See LICENSE.