Next week, delegates from 194 member states convene in Geneva for the seventy-ninth World Health Assembly. Health systems strengthening, universal health coverage, and the financing that connects the two will sit near the top of the agenda. Much of the evidence the room will draw on lives in two of WHO's flagship public databases — the Global Health Observatory (GHO) and the Global Health Expenditure Database (GHED). Today, we are open-sourcing two small tools that make those databases directly accessible from the AI assistants global-health researchers are already using every day.
Where WHO's data meets the AI workflow
The WHO Global Health Observatory and the Global Health Expenditure Database are two of the most important public goods in global health. The GHO publishes more than 2,500 indicators through a documented OData API; the GHED releases a comprehensive “all-data” workbook covering more than 4,000 variables across the System of Health Accounts 2011 framework. The teams at WHO who maintain both have built and curated the backbone of comparative health-systems research worldwide.
Researchers are also, increasingly, working with AI assistants. Claude, ChatGPT, Gemini, and others are part of the daily workflow in ministries of health, multilateral agencies, academic departments, and journalism teams across Latin America, the Caribbean, and beyond. A question gets asked in plain English, and the assistant either answers from training data — sometimes accurately, sometimes not — or runs an ad-hoc web search that does not always land on the most authoritative source.
What has been missing is a direct bridge between the two. AI assistants do not natively speak OData filter syntax, do not carry indicator codes like WHOSIS_000001 in their heads, and do not know that current health expenditure in GHED decomposes additively as CHE = HF1 + HF2 + HF3 + HF4 + HFnec under SHA 2011. With those primitives wired up, an assistant can do for a researcher what the researcher would do themselves — only faster, in any language, and without the schema lookup tax. That bridge is what gho-mcp and ghed-mcp provide.
What MCPs change
The Model Context Protocol, released by Anthropic in late 2024 and now an open standard, lets AI assistants like Claude talk to external systems through a uniform interface. The protocol defines how an assistant discovers what a tool can do, calls it, and renders the result. In practice, MCPs are how a model gains real, accountable access to anything outside its training data — a database, an API, a filesystem, a research workbook.
For a researcher, the relevant thing is this: when an MCP server is connected, you can ask a question in plain English and the model handles the schema, the codes, the loops, and the disaggregations on your behalf. The “Claude knows the answer” failure mode — confidently hallucinated numbers — disappears. The model is now reading the database; the database is the source of truth.
The shift in one line
Before: the model guesses, and you verify. After: the model queries, and you analyze.
gho-mcp: the Global Health Observatory, in plain English
gho-mcp is an MCP server for the WHO Global Health Observatory. It exposes a small set of task-shaped tools — country profile, compare countries, search indicators, describe indicator dimensions, resolve country codes, list curated country groups, and an expert-mode escape hatch for raw OData queries. Country names, ISO3 codes, WHO region codes (AFR, AMR, EMR, EUR, SEAR, WPR), and World Bank income groups (WB_HI, WB_UMI, WB_LMI, WB_LI) are all accepted interchangeably. CSV export is built in.
With it installed, an assistant can answer questions like:
- “Build me a comparative health-systems profile for Peru.”
- “Plot maternal mortality across the Andean countries since 2010.”
- “What is the UHC service coverage gradient by World Bank income group?”
- “Does the tobacco indicator support sex disaggregation?”
A country profile pulls roughly a dozen headline indicators in one parallel-fetched call — life expectancy, HALE, infant and under-five mortality, MMR, UHC service coverage, DTP3 coverage, skilled birth attendance, physician and nursing density, premature NCD mortality, tuberculosis incidence — all at the latest available year. The same shape works for any country in the GHO.
ghed-mcp: the Global Health Expenditure Database, accounting-aware
ghed-mcp is the companion server for the Global Health Expenditure Database. It discovers WHO's current “all-data” workbook on first use, builds a derived SQLite cache, and exposes a broader set of tools that respect how health-financing researchers actually work. The server knows the SHA 2011 additive hierarchies and validates breakdowns with a sum-versus-parent balance check, so a model is steered away from arithmetic that does not compute as an accounting identity.
With it installed, an assistant can answer:
- “Build me a health-financing profile for Colombia.”
- “Compare out-of-pocket burden across LAC countries since 2000.”
- “What is the government priority gradient by World Bank income group?”
- “Decompose Peru's current health expenditure by financing scheme for 2023.”
Beyond data extraction, there are research-workflow tools — data_availability, build_research_panel, build_research_package — that turn a vague question into a tidy long panel plus codebook and availability CSVs, ready to import into Stata or R.
A small example
Once installed, the surface is just natural language. A typical session in Claude might begin with:
Compare out-of-pocket health expenditure as a share of current health expenditure across LAC countries since 2000, and tell me which countries have moved furthest in the right direction.
Behind the scenes, the model calls compare_countries(indicator_code="oops_che", country_group="LAC", year_start=2000), computes percent change and CAGR with rank_country_changes, and writes back a paragraph naming the leaders and laggards together with a CSV ready to paste into a chart. No code is written by the researcher. The figures shown are the figures that exist in the database, with mixed-year warnings flagged where relevant.
A small choice with regional consequences
Both servers ship with a curated set of country groupings beyond what WHO and the World Bank publish as built-in dimensions. The most consequential of these is LAC — 33 sovereign Latin American and Caribbean states, the PAHO/Decilion convention. The World Bank's LAC region is 42 economies, which includes territories like Aruba, the Cayman Islands, Curaçao, and Puerto Rico that most LAC comparative work treats separately. By default, our LAC group is the 33-state one. The 42-economy version is also there, as LAC_TERRITORIES. So are OECD, LDC, SSA, MENA, and several others.
This is a small choice but a deliberate one. Regional aggregates shape every comparative health-systems argument, and which countries are inside the bracket matters. Researchers who want one definition or the other can pick. Researchers who don't know there is a choice still get the more analytically defensible default.
Install and what is next
Both servers are MIT-licensed and live on GitHub:
Installation is a short walk if you already have Python 3.11 — clone, create a virtualenv, pip install -e ., and register the server with your MCP client (Claude Code, Claude Desktop, Codex, Cursor, Cline, Continue). Both READMEs include the exact commands.
The servers are not yet listed on any MCP registry; that will come once the surface is stable. What helps right now is feedback — open an issue if a tool signature feels wrong, if a curated grouping needs a member added, or if there is a comparative question you cannot get either server to answer cleanly. The country lists are versioned and refreshed annually; Bangladesh, Lao PDR, and Nepal graduate from LDC status on November 24, 2026, and the file knows this.
The WHA convenes next week. By then, our hope is that a few dozen researchers, ministry staff, and journalists are looking at the same evidence the room is looking at, on their own questions, in their own languages. That is what we built these tools to do.
