{
  "markdown": "# 🔎 data-aggregator-mcp\n\n**One MCP server to find and fetch research data across archives, omics\nregistries, and literature — behind a single normalized model.**\n\n[![PyPI](https://img.shields.io/pypi/v/data-aggregator-mcp.svg)](https://pypi.org/project/data-aggregator-mcp/)\n[![Python](https://img.shields.io/pypi/pyversions/data-aggregator-mcp.svg)](https://pypi.org/project/data-aggregator-mcp/)\n[![Downloads](https://img.shields.io/pypi/dm/data-aggregator-mcp.svg)](https://pypi.org/project/data-aggregator-mcp/)\n[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://github.com/musharna/data-aggregator-mcp/blob/main/LICENSE)\n[![CI](https://github.com/musharna/data-aggregator-mcp/actions/workflows/ci.yml/badge.svg)](https://github.com/musharna/data-aggregator-mcp/actions/workflows/ci.yml)\n[![Glama](https://glama.ai/mcp/servers/musharna/data-aggregator-mcp/badges/score.svg)](https://glama.ai/mcp/servers/musharna/data-aggregator-mcp)\n[![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.21636332.svg)](https://doi.org/10.5281/zenodo.21636332)\n\n`search` one query across **17 sources** — **Zenodo, DataCite** (Dryad /\nFigshare / Dataverse / OSF / OpenNeuro / Mendeley), **NCBI omics**\n(GEO / SRA / BioProject), **BioStudies** (EBI, incl. ArrayExpress),\n**literature** (PubMed / OpenAIRE), **HuggingFace** datasets, **DataONE**\n(eco / environmental), **OmicsDI** (proteomics / metabolomics), **DANDI**\n(neurophysiology), **CZ CELLxGENE** (single-cell), **OpenML** (ML datasets),\n**RCSB PDB** (structures), **UniProtKB** (proteins), the **GWAS Catalog**,\n**GBIF** (biodiversity), **data.gov** (US federal open data), and **NASA CMR**\n(Earth science) — deduplicated, normalized, and cross-linked. `resolve` any hit to its file\nmanifest, citation, trust signals, and the data it points at. `fetch` it to\ndisk with checksum verification.\n\nmcp-name: io.github.musharna/data-aggregator-mcp\n\n<p align=\"center\">\n  <img src=\"https://raw.githubusercontent.com/musharna/data-aggregator-mcp/main/examples/assets/demo.svg\"\n       alt=\"data-aggregator-mcp stdio demo — initialize, tools/list (search, resolve, fetch, operate, relate, list_sources), and a live list_sources call showing the wired sources across archives, omics, and literature\"\n       width=\"820\">\n</p>\n\n## ✨ Why this\n\nMost data MCPs wrap a single source. This one **unifies** them behind six tools\nand one `DataResource` model, so an agent searches once and gets back comparable\nrecords:\n\n- **Multi-domain, one model** — generalist archives + raw omics + literature,\n  deduplicated by DOI (the fetchable record wins over bare metadata).\n- **Taxonomy synonym expansion** — `organism=\"Orobanche aegyptiaca\"` also matches\n  `Phelipanche aegyptiaca` (NCBI Taxonomy), so a species rename doesn't cost you\n  results.\n- **Paper → data bridge** — resolve a paper and get links to the GEO / SRA /\n  BioProject / DataCite records it produced.\n- **Verified fetch** — streams to disk with md5 verification where the source\n  exposes a checksum, optional archive unpacking, and a fail-loud integrity\n  sniff that rejects an HTML paywall page served as a \"PDF\".\n- **Citations, access & full text** — render a citation in any CSL style, get\n  normalized access/license, and pull open-access full text — all in one\n  `resolve`.\n- **Trust signals** — usage `metrics` (citations / views / downloads / likes),\n  version status (`is_latest` / `superseded_by`), and `last_updated` freshness,\n  surfaced wherever the source exposes them.\n- **Interop exports** — `resolve(format=\"croissant\")` or `\"ro-crate\"` hands a\n  dataset to an ML or research-packaging pipeline as standard JSON-LD.\n- **Operate on data in place** — `operate` reads the schema, previews rows, or\n  runs a read-only SQL `SELECT` against a remote Parquet/CSV/TSV **without\n  downloading it** (Parquet footer + DuckDB httpfs range reads). Optional\n  `[operate]` extra; base install is unchanged.\n- **Relate across records** — `relate` takes a handful of resolved ids and\n  reports how they connect — shared accession, shared cross-identifier, an\n  explicit link, or version lineage — naming the literal shared value as\n  evidence. Metadata hints only: it never reads files or executes a join.\n\n→ Full rationale and a comparison vs. single-source servers, breadth gateways, and\nML-dataset tools: **[docs/POSITIONING.md](https://github.com/musharna/data-aggregator-mcp/blob/main/docs/POSITIONING.md)**.\n\n<p align=\"center\">\n  <img src=\"https://raw.githubusercontent.com/musharna/data-aggregator-mcp/main/docs/assets/architecture.svg\"\n       alt=\"Architecture: an MCP client speaks stdio to data-aggregator-mcp's six tools, which fan out through one router (DOI dedup, ontology expansion, ranking) to archives (Zenodo, DataCite, HuggingFace, DataONE, OpenML, RCSB PDB), omics (GEO, SRA, BioProject, OmicsDI, DANDI, CELLxGENE, GWAS Catalog), and literature (PubMed, OpenAIRE, EuropePMC, Unpaywall)\"\n       width=\"760\">\n</p>\n\n## ⚡ Quickstart\n\nRun with no install:\n\n```bash\nuvx data-aggregator-mcp\n```\n\nRegister with Claude Code:\n\n```bash\nclaude mcp add data-aggregator -- uvx data-aggregator-mcp\n```\n\nA typical agent flow:\n\n```text\nsearch(\"drought stress RNA-seq\", organism=\"Sorghum bicolor\")\n  → [ geo:GSE..., sra:SRX..., zenodo:..., pubmed:... ]   # deduped, taxa-normalized\n\nresolve(\"sra:SRX079566\")\n  → DataResource{ files: [ENA FASTQ urls…], access: \"open\", taxa: [...] }\n\nfetch(\"sra:SRX079566\", dest=\"./data\")\n  → [\"./data/SRX079566_1.fastq.gz\", …]                   # md5-verified\n```\n\n<details>\n<summary>Other ways to run (pip, python -m, raw client config)</summary>\n\n```bash\npip install data-aggregator-mcp\ndata-aggregator-mcp        # or: python -m data_aggregator_mcp\n```\n\nTo use the `operate` tool (query remote tabular files in place), install the\noptional extra:\n\n```bash\npip install \"data-aggregator-mcp[operate]\"\n```\n\nAdd to a client's MCP config (e.g. Claude Desktop `claude_desktop_config.json`):\n\n```json\n{\n  \"mcpServers\": {\n    \"data-aggregator\": {\n      \"command\": \"uvx\",\n      \"args\": [\"data-aggregator-mcp\"],\n      \"env\": { \"NCBI_API_KEY\": \"your-optional-key\" }\n    }\n  }\n}\n```\n\n</details>\n\n## 🌐 Transports\n\n**stdio (default)** — the server runs as a child of the client, so `fetch()`\nwrites to your own disk. Nothing to configure; every command above uses it.\n\n**Streamable HTTP** — the same six tools, prompts, and resources over HTTP:\n\n```bash\ndata-aggregator-mcp --transport http     # → http://127.0.0.1:8000/mcp/\n```\n\n| flag                       | default          | notes                                                             |\n| -------------------------- | ---------------- | ----------------------------------------------------------------- |\n| `--transport {stdio,http}` | `stdio`          |                                                                   |\n| `--host`                   | `127.0.0.1`      | this machine only; any non-loopback value requires `--allow-host` |\n| `--port`                   | `8000`           |                                                                   |\n| `--allow-host HOST:PORT`   | auto on loopback | permitted `Host` header, repeatable — **required off loopback**   |\n| `--allow-origin ORIGIN`    | derived          | permitted browser `Origin` header, repeatable                     |\n| `--stateless`              | off              | fresh transport per request, no session affinity                  |\n| `--json-response`          | off              | plain JSON responses instead of SSE streams                       |\n\nThe endpoint is served at **`/mcp/`** — with the trailing slash. `/mcp` answers\n`307` redirecting there, which is fine for any client that follows redirects (a\n`307` preserves the POST body); point one that doesn't straight at `/mcp/`. In\nstateful mode, sessions idle for 30 minutes are reaped.\n\n**DNS-rebinding protection is always on.** A loopback bind derives its own\nhost/origin allowlist, so the default needs no configuration. A non-loopback bind\n(`--host 0.0.0.0`, a LAN address, a container interface) **refuses to start**\nwithout at least one explicit `--allow-host` — guessing an allowlist there is\nprecisely the hole the protection exists to close, so it fails loud instead of\nopen:\n\n```bash\ndata-aggregator-mcp --transport http --host 0.0.0.0 \\\n  --allow-host data.example.org:8000\n```\n\nOnce running, a request whose `Host` header is outside the allowlist is refused\nwith `421 Invalid Host header`.\n\n> ⚠️ **`fetch(dest=…)` writes to the _server's_ filesystem, not the client's.**\n> Over stdio those are the same disk; over HTTP they may be different machines,\n> and the caller gets back paths it cannot read. Treat `dest` on an HTTP\n> deployment as server-side staging, or use stdio when you need the bytes\n> locally. `search`, `resolve`, `operate`, `relate`, and `list_sources` are\n> unaffected — they return data, not paths.\n\n## 🗂️ Sources\n\n| Source                       | Discover |       Fetch       |     Checksum     |\n| ---------------------------- | :------: | :---------------: | :--------------: |\n| Zenodo                       |    ✅    |        ✅         |       md5        |\n| DataCite → Figshare          |    ✅    |        ✅         |       md5        |\n| DataCite → Dataverse         |    ✅    |        ✅         |       md5        |\n| DataCite → OSF               |    ✅    |        ✅         |       md5        |\n| DataCite → Dryad             |    ✅    |  manifest only¹   | sha-256 (listed) |\n| DataCite → Mendeley & others |    ✅    |         —         |        —         |\n| NCBI SRA                     |    ✅    |  ✅ (ENA FASTQ)   |       md5        |\n| NCBI GEO                     |    ✅    |   ✅ (`suppl/`)   |      none²       |\n| NCBI BioProject              |    ✅    |    → SRA links    |        —         |\n| PubMed / OpenAIRE            |    ✅    | ✅ (OA full text) |      none²       |\n| HuggingFace datasets         |    ✅    | ✅ (resolve URL)  |       none       |\n| DataONE (eco/env)            |    ✅    | ✅ (Member Node)  |  md5 / sha-256   |\n| OmicsDI → PRIDE              |    ✅    |  ✅ (HTTPS FTP)   |    size only     |\n| OmicsDI → MetaboLights       |    ✅    |  ✅ (HTTPS FTP)   |       none       |\n| OmicsDI → other MS repos     |    ✅    |         —         |        —         |\n| DataCite → OpenNeuro         |    ✅    |   ✅ (snapshot)   |      none²       |\n| DANDI (neurophysiology)      |    ✅    |    ✅ (302→S3)    |      none²       |\n| CZ CELLxGENE (single-cell)   |    ✅    |   ✅ (H5AD/RDS)   |      none²       |\n| OpenML (ML datasets)         |    ✅    |     ✅ (ARFF)     |       md5        |\n| RCSB PDB (structures)        |    ✅    |  ✅ (.cif/.pdb)   |      none²       |\n| GWAS Catalog                 |    ✅    |   → PMID bridge   |        —         |\n\n¹ Dryad downloads are token / bot-challenge gated, so `fetch` fails loud;\n`resolve` still lists the files.\n² No upstream checksum — `fetch` verifies content-type instead (rejects an HTML\npage served in place of a binary).\n\n## 🛠️ Tools\n\n### `search(query?, size?, sources?, organism?, disease?, tissue?, chemical?, assay?, kind?, published_after?, published_before?, rank?, cursor?, collapse_mirrors?, understand?, multi_query?, provenance?)`\n\nFan out across all wired sources in parallel and return compact `DataResource`\nrecords, deduped by DOI. Per-source failures land in `errors{}` — never silently\ndropped.\n\n- `organism` — expand the query with NCBI-Taxonomy synonyms; the expansion is\n  echoed in `taxon_expansion`, and results carry normalized `taxa[]`\n  (`{taxid, name}`) plus a `described_in` link to plant-genomics-mcp for plant\n  taxa.\n- `sources` — restrict the fan-out, e.g. `[\"omics\"]`.\n- `size` — max results (1–50).\n- `kind` — keep only `dataset` / `sequencing_run` / `study` / `publication` /\n  `software`.\n- `published_after` / `published_before` — filter by publication year.\n- `rank` — `relevance` (default) or `semantic` (re-rank the fetched page by\n  embedding similarity to the query; needs `EMBEDDING_API_BASE`, degrades to\n  relevance order otherwise).\n- `understand` — opt into LLM query understanding (default false). A free-text\n  query is **normalized** into a focused keyword query: conversational fluff\n  (`\"I'm looking for…\"`, `\"where can I find…\"`) is stripped while the scientific\n  and entity terms are kept so they still match by text. The LLM also detects\n  structured entities (organism/disease/tissue/chemical/assay, kind) — these are\n  **echoed in `query_understanding.extracted` for transparency but not\n  auto-applied**, because ANDing LLM-_inferred_ facets across free-text keyword\n  upstreams over-constrains and hurts recall. Only the cleaned `keyword_core` and\n  explicit `year` scopes are applied; the ontology resolvers still run on the\n  facets **you** pass (the LLM proposes, you dispose). Needs an LLM endpoint\n  (`LLM_API_BASE`); with none configured the search runs unchanged and notes it in\n  `errors['understand']`. **Effectiveness is query- and model-dependent — opt-in /\n  default-off; validate the recall lift on your own corpus and LLM (see the eval\n  harness below). On our small verified set `multi_query=` is the stronger,\n  always-safe recall lever; `understand=` is approximately neutral with a weak\n  local model.**\n- `multi_query` — opt into diverse multi-query recall expansion (default false).\n  An LLM generates up to a few deliberately-diverse reformulations of your query\n  (different facets/synonyms/framings, not paraphrases), each is fanned out across\n  every source, and the deduped union is re-ranked against your **original** query —\n  surfacing relevant records a single keyword query would miss. Bounded at\n  `MAX_QUERY_VARIANTS` (4, incl. the original, which is always kept so recall never\n  drops below baseline), so it costs at most N× the upstream calls. Composes with\n  `understand=` (which structures variant 0). The variants used are echoed in\n  `query_expansion`. Needs an LLM endpoint (`LLM_API_BASE`); with none configured\n  the search runs as a normal single query and notes it in `errors['multi_query']`.\n- `cursor` — opaque token from a prior result's `next_cursor`; pages forward\n  across every source. In `cursor` mode the other params are read from the\n  token, so `query` is optional.\n\n### `resolve(id, cite?, format?, trust?, fair?, use?)`\n\nFull record + files manifest. Routes by id shape — `zenodo:7654321`, a bare DOI,\n`datacite:10.5061/dryad.x`, an omics id (`sra:SRX079566`, `geo:GSE332789`,\n`bioproject:PRJNA1468572`), a literature id (`pubmed:34320281`, `openaire:<id>`),\na HuggingFace id (`hf:owner/name`), a DataONE id (`dataone:doi:10.5063/F1HT2M7Q`),\nor an OmicsDI id (`omicsdi:pride:PXD000001`). Attaches, where available:\n\n- **`files[]`** — ENA FASTQ manifest (SRA), GEO `suppl/`, or the host repo's\n  native manifest (Figshare / Dataverse / OSF / Dryad).\n- **`links[]`** — paper → data: `pubmed:` → `sra:` / `geo:` / `bioproject:` (NCBI\n  elink); `openaire:` → `datacite:` (ScholeXplorer Scholix).\n- **`access` / `license`** — normalized status\n  (`open` / `embargoed` / `restricted` / `closed` / `unknown`) and license where\n  the source exposes it.\n- **`identifiers`** — normalized `{pmid, pmcid, doi}`, plus an open-access\n  full-text `FileEntry` (EuropePMC XML, or an Unpaywall PDF fallback) for papers.\n- **`citation`** — pass `cite=<format>`: `bibtex`, `ris`, `csl-json`, or any CSL\n  style name (`apa`, `mla`, `vancouver`, …). DOI records use content\n  negotiation; others render CSL-JSON from metadata. Off by default; failures\n  degrade quietly.\n- **trust signals** — `metrics` (citations / views / downloads / likes),\n  `is_latest` / `superseded_by` (derived from version links), and `last_updated`\n  freshness, where the source provides them.\n- **`trust=true`** — attach retraction status (via Crossref) under `trust{}`.\n  One extra Crossref call; meaningful for DOI-bearing records only.\n- **`fair=true`** — attach an RDA-grounded FAIRness score (0–100 + F/A/I/R\n  sub-scores + actionable gaps) computed from the record metadata under `fair{}`.\n  Pure/local — no extra network call.\n- **`use=<intent>`** — attach a licence-compatibility advisory under\n  `license_compat{}` for the intended use (`commercial` / `redistribute` /\n  `modify` / `ml-training`). Returns ALLOW/REVIEW/DENY with the governing clause.\n  Metadata-derived advisory, **not legal advice**; an absent/unrecognized licence\n  yields REVIEW.\n- **`format`** — pass `format=\"croissant\"` (file-level Croissant JSON-LD),\n  `\"ro-crate\"` (minimal RO-Crate 1.1), or `\"provenance\"` (one-call RO-Crate 1.1\n  data-availability dossier bundling version-currency, licence+SPDX, FAIR score,\n  and retraction status) to attach a standard manifest under the matching field.\n\n### `fetch(id, dest?, files?, max_bytes?, force?, extract?)`\n\nDownload files to disk and return their paths. Streams under a `max_bytes` guard\n(`force` to override) with md5 verification wherever a checksum exists.\n\n- `files` — restrict to a subset of the resolved manifest.\n- `extract` — unpack downloaded zip / tar archives in place, guarded against\n  path traversal and runaway extracted size. Off by default.\n- Unverified fetches (GEO `suppl/`, literature full text) get a content-type\n  sniff that fails loud if a declared binary is actually an HTML page.\n- Fetchable: **Zenodo**, **SRA**, **GEO**, **DataONE** (Member-Node objects,\n  md5/sha-256 verified), DataCite-hosted **Figshare** / **Dataverse** / **OSF**,\n  **HuggingFace** datasets, **PRIDE** / **MetaboLights** (via OmicsDI, unverified),\n  and **literature** open-access full text. **Dryad**, other DataCite repos, and\n  other OmicsDI repos (MassIVE / GNPS / ...) are discovery-only and raise\n  `FetchNotSupportedError`.\n\n### `list_sources()`\n\nWired sources with their capabilities — layer, kinds, supported filters,\nfetchability, `operable` flag, id examples, auth, and rate limits.\n\n### `operate(op, id, file?, query?, n?, columns?)`\n\nInspect or query a remote tabular file (Parquet / CSV / TSV) **without\ndownloading it**. Addresses a file by catalog `id` + `file` name (defaults to the\nfirst tabular file on the resolved record). Ops:\n\n- `schema` — column names + types (reads the Parquet footer / sniffs the CSV\n  header; no full load).\n- `preview` — a small sample of rows.\n- `head` — the first `n` rows (default 20), optionally restricted to `columns`.\n- `sql` — a read-only `SELECT` (the file is the view `data`), e.g.\n  `SELECT col, count(*) FROM data GROUP BY 1`.\n- `peek` — per-column profile via DuckDB `SUMMARIZE` (type, null-rate,\n  approximate distinct count, min/max, numeric quartiles) **without\n  downloading** the file. Like `head`/`sql`, reads the whole file and honors\n  the source-size ceiling.\n\nBacked by the Parquet footer reader + DuckDB `httpfs` range reads. `sql` runs in\na locked-down DuckDB (read-only, local filesystem disabled, single-SELECT\nvalidation, row / wall-clock caps). Requires the optional `[operate]` extra\n(`pip install data-aggregator-mcp[operate]`); without it, `operate` returns a\nclear install-the-extra message and the other four tools are unaffected.\n\nAny HuggingFace dataset with a datasets-server converted view is operable\n(`schema` / `preview` / `head` / `sql`): `resolve` surfaces the auto-converted\nParquet files (`source=\"hf-datasets-server\"`) even for datasets stored as\nJSON/JSONL/arrow, so pass `file=<config>/<split>/...parquet` to pick a split when\nthere are several.\n\n### `relate(ids)`\n\nCross-resource join/harmonization **hints**. Given 2–10 resource ids, `relate` resolves\neach (TTL-cached) and reports how they relate and on what key they could be joined:\n\n- **`shared_accession`** — same BioProject/SRA/GEO accession on ≥2 records → joinable key.\n- **`shared_identifier`** — same doi/pmid/pmcid across records → same work / paper↔data link.\n- **`explicit_link`** — one record's `links[]` points at another input record.\n- **`version_lineage`** — one record supersedes another (dedupe, don't join, those).\n\n**Hints only.** `relate` never reads file columns, fetches files, or executes a\njoin/merge/conversion — every hint names the shared value as evidence. Per-id resolve\nfailures are reported in `errors`, not fatal; an empty result carries an explanatory\n`note`.\n\n### Prompts\n\nThree workflow prompts surface in clients (e.g. `/mcp__data_aggregator__*` in\nClaude Code):\n\n- **`find_data`** — find datasets for a topic, optionally scoped to an organism.\n- **`data_behind_paper`** — find the datasets / accessions behind a paper.\n- **`search_resolve_fetch`** — walk the end-to-end search → resolve → fetch flow.\n\n## ⚙️ Configuration\n\nAll optional, set via environment variables:\n\n- `NCBI_API_KEY` — raises the NCBI E-utilities rate limit (3 → 10 req/s) used by\n  the omics, literature, and taxonomy lookups.\n- `DATA_GOV_API_KEY` — a free [api.data.gov](https://api.data.gov/signup/) key for\n  the data.gov source. Absent ⇒ requests fall back to the shared public\n  `DEMO_KEY`, rate-limited to roughly 30 requests/hour per IP — fine for light\n  discovery, worth setting for anything heavier.\n- `UNPAYWALL_EMAIL` — enables the Unpaywall fallback leg of literature full-text\n  retrieval (the EuropePMC leg works without it).\n- `NCBI_EMAIL` — contact address sent to NCBI's ID converter; falls back to\n  `UNPAYWALL_EMAIL` when unset.\n- `DATAVERSE_BASE_URL` — resolve Dataverse DOIs against a different installation\n  (default `https://dataverse.harvard.edu`).\n- `CACHE_TTL_SECONDS` — resolve-cache lifetime in seconds (default `3600`; an\n  unparseable value falls back to that default).\n- `EMBEDDING_API_BASE` / `EMBEDDING_API_KEY` / `EMBEDDING_MODEL` — an\n  OpenAI-compatible embeddings endpoint enabling `rank=semantic`. Absent ⇒\n  semantic re-rank degrades to relevance order. Key is optional (keyless local\n  servers supported); model defaults to `text-embedding-3-small`.\n- `LLM_API_BASE` / `LLM_API_KEY` / `LLM_MODEL` — an OpenAI-compatible\n  `/chat/completions` endpoint enabling `search(understand=true)` (NL→structured\n  query rewriting) **and** `search(multi_query=true)` (diverse multi-query recall\n  expansion). Absent ⇒ both run the raw query unchanged and note it in\n  `errors['understand']` / `errors['multi_query']`. Key is optional (keyless local\n  servers supported); model defaults to `gpt-4o-mini` (a passthrough string — set\n  it to whatever your endpoint serves). `multi_query` fans out at most\n  `MAX_QUERY_VARIANTS` (4, incl. the original) variants, bounding the N× cost.\n\nTo measure the recall lift of `understand=true` / `multi_query=true` on a small\nlabeled set, run the gated eval harnesses (need a live LLM endpoint):\n\n```bash\nDATA_AGGREGATOR_MCP_LIVE=1 LLM_API_BASE=... python scripts/eval_understand.py\nDATA_AGGREGATOR_MCP_LIVE=1 LLM_API_BASE=... python scripts/eval_multi_query.py\n```\n\nThey print per-query and mean recall@20 (understand / multi-query off vs. on). See\nthe fixtures at `scripts/eval_understand_fixture.json` and\n`scripts/eval_multi_query_fixture.json`.\n\n## 🧪 Develop\n\n```bash\nuv venv && uv pip install -e \".[dev]\"\nuv run pytest -q\nuv run ruff check src tests\nDATA_AGGREGATOR_MCP_LIVE=1 uv run pytest -k live -q   # real-API probes\n```\n\nThe README demo (`examples/assets/demo.svg`) is recorded network-free from\n`examples/_demo_stdio.py` — see the header of that file to re-record.\n\n## License\n\nMIT — see [LICENSE](https://github.com/musharna/data-aggregator-mcp/blob/main/LICENSE).\n",
  "bytes": 23369,
  "sha": "352f76b2cae062b1f72a1710a17fe68c572249b34ed13cd5a2fa2846edeec57d",
  "repo_slug": "musharna/data-aggregator-mcp",
  "fonte": "repo",
  "truncated": false,
  "api": "https://agentalog.com/api/listings/mcp_io_github_musharna_data_aggregator_mcp_a711aad9/readme"
}