UNESCO UIS — Education, Science & Culture Statistics (provenance-first)

UNESCO Institute for Statistics (UIS) — MCP Server

[MCP] [CI] [Version] [Tools] [npm] [MCP Registry] [uis-mcp-server MCP server] [smithery badge] [License: MIT] [Status]

🇧🇷 Leia em Português

A public, hosted, provenance-first MCP server for the statistics of the UNESCO Institute for Statistics (UIS) — education, science and R&D, culture and communication — no installation, no account, no API key. Point your MCP client at the hosted endpoint and ask about enrolment, completion, out-of-school children, literacy, education spending, researchers and R&D expenditure by country, region and year. It runs on Cloudflare Workers over Streamable HTTP and talks to the official UIS Data API.

Independent project. This is an unofficial, community-built client of the UIS's public Data API — not affiliated with or endorsed by UNESCO or the UNESCO Institute for Statistics. Data remain © UNESCO-UIS under CC BY-SA 4.0; see Data license and attribution.

Every response carries a provenance block (source URL, the UIS data release it came from, real retrieval timestamp, license, UIS citation) — exact figures with an audit trail, not numbers guessed from training data.

🇧🇷 Em português. Servidor MCP remoto e hospedado (nada para instalar, sem conta e sem chave) para as estatísticas do Instituto de Estatística da UNESCO — matrícula, conclusão, crianças fora da escola, alfabetização, gasto em educação e P&D por país, região e ano, direto no Claude, no ChatGPT ou em qualquer cliente MCP, com proveniência e citação da fonte em cada resposta: README em português.

Questions it answers

In plain language, inside the MCP client — the assistant picks the tool and the filters:

• "Which UNESCO indicators track out-of-school children?" (uis_search_indicators)
• "How did primary completion evolve in Brazil, Peru and Colombia after 2015?" (uis_get_data)
• "Is there a UNESCO series on government expenditure on education as a share of GDP?" (uis_search_indicators → uis_get_data)
• "How many researchers per million inhabitants do the BRICS countries have since 2010?" (uis_search_indicators → uis_get_data)
• "What is the adult literacy rate in Argentina and Chile, by sex?" (uis_get_data)
• "Which geo unit code does the UIS use for Côte d'Ivoire, and which regional aggregates exist?" (uis_list_geo_units)

Ask in your words, not UNESCO's. The UIS is worded in British statistical English, and the catalogue is matched on the indicator name — so the everyday or US word used to return nothing at all. Measured over the 5,063 indicators of the official catalogue (2026-09-16), and fixed since 0.3.0: the search expands the term to the source's wording (OR within a term, AND across terms — it only adds recall) and says it translated, in vocabulary_notes; a zero result comes with a hint on what to try next. The same table feeds the search index (Deep Research).

Only measured pairs go in (the asked word absent from the catalogue, the source's word present) — the table lives in src/uis/vocabulary.ts, with the counts. A term the UIS does not publish is left out and still returns zero, because an alias for data that does not exist promises what the source does not have: dropout, tuition, unemployment and labor (labour statistics belong to the sister server ilo-mcp-server; what labor matches today are 62 names containing "collaboration").

Comparison with the alternatives

Anyone who already works with UIS data has good tools, and this server replaces none of them — it sits somewhere else in the chain: it answers the question at the point where the question is asked, inside the assistant, with source, data release and licence attached to the answer. Versions measured on 2026-10-02.

About SDMX: the UIS no longer serves it. Its legacy SDMX API reached end-of-life on 2020-06-23 (stated in the UIS's own February 2026 data release note) and the endpoint now answers 404 — generic SDMX clients that still list a UNESCO source (e.g. sdmx1) point at that dead URL, and the DBnomics copy of UNESCO data is a frozen 2020-era snapshot. For SDMX over labour statistics, see the sister server ilo-mcp-server, which talks to the ILOSTAT SDMX API.

Do not use this server when you need a whole dataset rather than an answer (the bulk files are the right tool), when the question is not education, science, culture or communication statistics published by the UIS (labour → ILO, national accounts → IMF/World Bank), or when you need microdata: the UIS publishes aggregates, and so does this server.

Sister servers, same design and same provenance block, for other official sources: ILOSTAT (ILO labour statistics), IBGE (Brazilian statistics), BCB (Central Bank of Brazil), Senado (Brazilian Senate open data), SIH/SUS (Brazilian hospital admissions) and medical terminologies (ICD-11, ICD-10, LOINC, RxNorm, ATC, MeSH).

Use it (hosted — no setup)

Point any MCP client at the Streamable HTTP endpoint:

https://uis.sidneybissoli.com/mcp

Claude Desktop / Claude Code and other clients with native remote support:

{  "mcpServers": {    "unesco-uis": {      "url": "https://uis.sidneybissoli.com/mcp"    }  }}

For clients that launch MCP servers as a command, use the mcp-remote bridge:

{  "mcpServers": {    "unesco-uis": {      "command": "npx",      "args": ["-y", "mcp-remote", "https://uis.sidneybissoli.com/mcp"]    }  }}

The uis-mcp-server.sidneybissoli.workers.dev hostname is also served, as a secondary.

ChatGPT (Deep Research)

ChatGPT deep research (and company knowledge, and research workflows over the Responses API) only uses an MCP server that exposes exactly search and fetch — this server does, on top of the uis_* tools. Point the connector at the hosted endpoint, no key required:

https://uis.sidneybissoli.com/mcp

search ranks the query against the full UIS catalogue (~5,060 indicators — education, science/R&D, culture, demographic context) and returns { id, title, url } (ind:<code>, e.g. ind:ROFST.1.CP); fetch returns the indicator as readable Markdown — name, theme, Data Browser group and framework, available years, a data sample (Brazil and the SDG world aggregate, last five years; one Data API call with a pinned release) and how to query it with uis_get_data — with the public UIS Data Browser page as url (https://databrowser.uis.unesco.org/view#indicatorPaths=<framework>%3A0%3A<code>), which is what ChatGPT cites. In search/fetch the text channel is the Deep Research contract's JSON (no footer); provenance travels in structuredContent and _meta. In ChatGPT's developer mode (Settings → Security and login → Developer mode) any tool is callable — the uis_* tools remain the ones to use for data.

Run locally (stdio)

Prefer not to route queries through a third-party host? The same server also runs as a local stdio process that talks directly to the official UIS Data API — same 5 tools and 3 resources, same limits, same provenance block, no Cloudflare in the loop. The surface of the two runtimes is identical by construction (CI checks it: node scripts/dump-surface.mjs --stdio).

No install needed — the package is on npm (uis-mcp-server, Node ≥ 22):

{  "mcpServers": {    "unesco-uis": {      "command": "npx",      "args": ["-y", "uis-mcp-server"]    }  }}

Or from source:

git clone https://github.com/SidneyBissoli/uis-mcp-servercd uis-mcp-servernpm installnpm run buildnode dist/cli.js   # serves MCP over stdio (Ctrl+C to stop)

Differences from the hosted server, all due to the absence of Cloudflare bindings:

• the current data release (/versions/default) is kept in process memory: resolved once per session, not across sessions;
• the indicator catalogue and the geo units are downloaded from the official endpoints on the first search (/definitions/indicators, /definitions/geounits — the same ones the hosted catalogue is seeded from; ~1 s), and the real retrieved_at of that download is what provenance reports;
• the in-memory catalogue does not download the UIS Data Browser definitions (~6.7 MB), so the URL that fetch cites falls back to the Data Browser home page and the search index lacks the group name in its keywords — Deep Research talks to the hosted server, which has them;
• no usage metrics, rate limit or auth.

Logs go to stderr — stdout carries only the JSON-RPC stream. The repository Dockerfile builds this runtime (used by the Glama registry).

Tools

Typical flow: uis_search_indicators to find an indicator code → uis_list_geo_units for country/region codes → uis_get_data with year filters.

Every response carries the portfolio's provenance block (@sbissoli/mcp-provenance, modes concise/detailed via the provenance_mode parameter) on three channels: structuredContent, namespaced _meta (com.sidneybissoli.uis/*) and a text footer. Since 1.1.0 the block includes retrieval, the origin diagnostic of the call — how many requests went to the Data API, how many attempts they took, which anomalies were overcome and whether the answer is unstable — measured by the portfolio's common fetch (@sbissoli/mcp-upstream); null when the answer did not touch the UIS (catalogue or cached release).

Resources

Three resources (static, text/markdown, no upstream call) that a client can attach to the context before calling tools — they save the discovery calls most sessions spend on "which indicator, which code":

Behaviour and limits

• Every data query pins the release. uis_get_data sends an explicit version=, resolved from /versions/default and cached for 24 h — reproducible, and it reuses the upstream's own cache. The release is the data_vintage (e.g. 20260507-91260335 (published 2026-05-08)); when the UIS publishes a new release, data queries move to it within 24 h.
• 5,000 records per response, at most 25 indicators per call. Above that the server returns an explanatory error with the real count — it never truncates silently, since partial data presented as complete would break the provenance contract. The UIS's own ceiling of 100,000 records comes back as the upstream's explanatory HTTP 400, passed through.
• The indicator catalogue is a snapshot of a UIS release, refreshed by hand when the UIS publishes a new one (2–3 times a year); its real retrieved_at is reported in the provenance of uis_search_indicators, so its age is always visible (served_from_cache: true).
• retrieved_at is always the real instant of extraction from the UIS, never the build or response time. When an answer joins parts extracted at different moments (uis_get_data: the current release, cached up to 24 h, plus the data fetched now; fetch: also the catalogue row), it is the OLDEST of them, and field_sources gives each part its own instant and cache flag (since 1.5.1).
• Notices report footnote types, magnitude and qualifier, with counts; the full text of each footnote stays on the row (include_footnotes: true).
• Every upstream call has a timeout and a retry policy (since 1.1.0, measured against the live API on 2026-09-27: release 1.1 s, small query 0.8 s, five whole indicators 2.9 s/3.5 MB, the 100k-cap 400 in 0.9 s): 35 s per attempt — ten times the worst legitimate case and above the 29 s of the AWS API Gateway in front of the UIS —, up to 3 attempts on 5xx, 429 (honouring Retry-After) and network errors, 45 s in total per call. Timeouts, HTTP 504, the explanatory 400 and 404 are never retried. What happened is reported in the provenance retrieval field; a timeout or network failure becomes a readable error, not a raw exception.
• derived is always false — the server does not transform values.
• Language: English; timezone: UTC (UIS data is published in English).

Surface fingerprint, checkable from the registry

Every release publishes, in its MCP Registry entry, the sha256 of the surface this version serves (initialize, tools, resources, prompts) and which methods answer without a credential. A surface change without a version bump fails the build; the fingerprint in the registry lets a client check the same thing on its own side. Canonical form and procedure: SPEC.md of @sbissoli/mcp-surface. To check it yourself (Node 18+, no dependencies):

curl -sO https://raw.githubusercontent.com/SidneyBissoli/mcp-br-commons/main/packages/mcp-surface/exemplos/verify.mjsnode verify.mjs io.github.SidneyBissoli/uis-mcp-server

Data license and attribution

• UIS data: CC BY-SA 4.0 (terms of the UIS Data Browser, which govern the Data API; license verified 2026-08-04).
• UIS attribution in every response (citation field), with the full URL and the extraction date: Source: UNESCO Institute for Statistics (UIS), <URL>, date of extraction <date>.
• ShareAlike: anything you redistribute that is built on these data must keep the CC BY-SA 4.0 license — the uis://reference/provenance resource spells it out.
• The UNESCO logo is not used and no endorsement is implied; that is why the server is called uis-mcp-server and not "unesco-mcp-server". The ILOSTAT data (CC BY 4.0) live in a separate server, so the two licences never share a response.

Self-hosting / development

Everything below is only needed to run your own instance — it is not required to use the public server.

npm installnpm run typecheck && npm test   # offline suite (tools, framework, vocabulary, in-memory catalogue, eval fixtures)npm run dev                     # http://localhost:8787/mcp (Worker)npm run build                   # stdio runtime → dist/cli.js (what ships to npm)node scripts/dump-surface.mjs --stdio   # offline smoke of the build: initialize → tools/resources/prompts
# Catalogue seed (D1) — required before first use:node scripts/seed-uis-catalog.mjsnpx wrangler d1 execute uis-catalog --local  --file=scripts/seed-uis-catalog.sqlnpx wrangler d1 execute uis-catalog --remote --file=scripts/seed-uis-catalog.sql
npm run deploynode scripts/smoke-mcp.mjs      # smoke test against production (initialize → tools/list == /status → uis_* → search/fetch → errors)

Notes for operators:

• Catalogue refresh is manual, once per UIS data release (no cron). The production smoke prints the current Data API release; when it differs from the seed's release, re-seed with the three commands above. Data queries follow the default release on their own, so only the catalogue can age — and its age is exposed in provenance.
• The UIS Data API accepts Node's fetch as is (no header workaround needed, unlike the ILO gateway).
• Optional Bearer auth (wrangler secret put API_KEY); token-bucket rate limit per IP.

Evals

@sbissoli/mcp-evals: 22 fixtures in evals/fixtures/queries.ts, validated offline in npm test. The run with a real model (npm run eval) uses the Anthropic API and needs ANTHROPIC_API_KEY (without it, it exits with instructions). Run of 2026-08-07: top-1 100% (20/20) — evals/results/.

End-to-end: 10 complex questions with a single verifiable answer in evals/e2e/evaluation.xml, answers validated manually against production (evals/e2e/validacao-respostas.md). Run of 2026-08-07 (Sonnet): 10/10 — evals/results/2026-08-07-e2e.md.

Endpoints

Security

Snyk Agent Scan (2026-08-07, before search and fetch were added): passed — report in security/.

License

Code: MIT. Data: UIS, CC BY-SA 4.0 (see "Data license and attribution" above).

Privacy

Privacy policy of the hosted service: PRIVACY.md.

Contact

Sidney da Silva Pereira Bissoli — [email protected]. This service is not endorsed by UNESCO.