unicefstats-mcp

Experimental — not an official UNICEF product. Verify retrieved values against the UNICEF Data Warehouse before citing in publications. See Limitations.

MCP server for UNICEF child development statistics. Query 790+ child-focused indicators across 200+ countries with disaggregations by sex, age, wealth quintile, and residence. No API key required.

Indicators cover child mortality, nutrition, education, child protection, WASH (water/sanitation/hygiene), HIV/AIDS, immunization, early childhood development, and more. Many align with SDG targets, but the dataset is broader than SDGs alone.

Data source: UNICEF SDMX API

Identity

Third-party aggregator listings (LobeHub, Smithery, mcp.so, Glama) are not controlled by the maintainer. Verify against the canonical source above.

Contents

• How it relates to the unicefdata packages

• How it compares to other data MCPs

• Landscape: MCP servers for official statistics

• Relationship to sdmx-mcp

• Quick Start

• Tools

• Demo

• Prompts

• Benchmark Results

• Deployment

• Development

• Contributing

• Limitations and Hallucination Risks

• Provenance and Ownership

• How to Verify This MCP

• License

Key documents

Related MCP server: MoSPI MCP Server

How it relates to the unicefdata packages

unicefstats-mcp is not a replacement for the unicefdata packages in Python, R, or Stata. They serve different audiences:

Under the hood, unicefstats-mcp wraps the unicefdata Python package. Every tool call ultimately calls unicefdata.unicefData() or its metadata functions. Think of the MCP as a thin AI-friendly interface on top of the same data layer.

When to use which:

• Use unicefstats-mcp when you're chatting with an AI and want to quickly explore indicators, check values, or compare countries

• Use unicefdata (Python/R/Stata) when you're writing scripts, building dashboards, running regressions, or doing any reproducible analytical work

How it compares to other data MCPs

Landscape: MCP servers for official statistics

This project is part of a growing ecosystem of MCP servers for international and official statistics. As of March 2026:

UN Agencies

International Organizations

National Statistics Offices

Known gaps

No MCP server exists for: FAO/FAOSTAT, UNESCO/UIS (4,000+ education indicators), ILO/ILOSTAT, UNSD SDG API, UN DESA Population, UNDP/HDI.

Full directory with install commands: MCP-DIRECTORY-STATS.md

Relationship to sdmx-mcp

UNICEF also maintains sdmx-mcp, a generic SDMX protocol MCP server. The two servers are complementary, not competing:

Key tradeoff: unicefstats-mcp is dramatically more accurate (EQA 0.891 vs 0.074) because its formatted output is optimized for LLM parsing. sdmx-mcp achieves zero hallucination on absent-data queries through aggressive assistant_guidance fields and a validate_query_scope pattern; its accuracy floor is too low to be useful, but the refusal discipline is exemplary. unicefstats-mcp v0.7.3 + fixes is the first version where MCP demonstrably makes the model safer than the no-tools baseline on absent-data queries (hall_b < hall_a), achieved without sdmx-mcp's accuracy cost.

When to use which:

• Use unicefstats-mcp for UNICEF child development analysis — it's simpler, faster, and far more accurate

• Use sdmx-mcp when you need to query non-UNICEF SDMX registries, explore dataflow structures, or work with hierarchical codelists

Full 3-way benchmark (LLM alone vs unicefstats-mcp vs sdmx-mcp): examples/results/

Quick Start

pip install unicefstats-mcp

Claude Code

Add to ~/.claude/.mcp.json:

{
  "mcpServers": {
    "unicefstats": {
      "command": "unicefstats-mcp"
    }
  }
}

Cursor / VS Code

Add to your MCP settings:

{
  "unicefstats": {
    "command": "unicefstats-mcp"
  }
}

What v1.1.0 adds

v1.1.0 is additive: v1.0.0's ambiguity_flag, candidates, and abstain_instruction still fire unchanged on ambiguous queries. v1.1.0 layers four new advisory envelope fields on top:

• requires_confirmation — true when the server wants the assistant to pause for user input

• recommended — the server's preferred next action (code, country, year, or tool)

• assistant_guidance — short natural-language hint the LLM can paraphrase

• next_step — a structured tool/parameter suggestion the LLM can chain into

Decision order (4 stages): strict canonical lookup → ambiguity check (v1.0.0 fields) → confirmation gate (requires_confirmation) → advisory hints (recommended / assistant_guidance / next_step).

Verdict: ALLOW (no behavioural regression; advisory fields fire on 23/30 Sonnet and 26/30 Haiku paired stuck queries; see internal/v1.1.0_design/ab_results.md).

Tools

Workflow

1. search_indicators("child mortality")     → find indicator codes
2. get_indicator_info("CME_MRY0T4")         → check disaggregations & SDMX details
3. get_temporal_coverage("CME_MRY0T4")      → check year range
4. get_data("CME_MRY0T4", ["BRA", "IND"])   → fetch data
5. get_api_reference("python", "unicefData") → get code template to continue in a script

Resources

The server exposes six MCP resources clients can load for guidance and reference data:

The system-prompt and context resources address the T2 hallucination failure mode (model fabricating values for years beyond the data frontier). Pattern adopted from the World Bank data360-mcp server. See CHANGELOG entry for v0.5.0.

Scope: UNICEF DW indicators only

unicefstats-mcp searches and serves the UNICEF Data Warehouse SDMX catalog — approximately 790 child-focused indicators across mortality, nutrition, education, child protection, WASH, HIV/AIDS, immunization, and early childhood development. Indicator codes follow UNICEF conventions (e.g. CME_MRY0T4, NT_ANT_HAZ_NE2, ED_ANAR_L1).

Codes from other organisations are out of scope. The MCP cannot find, resolve, or fetch them because they do not exist in the UNICEF SDMX catalog. Examples of out-of-scope code families:

These codes will return no hits from search_indicators and a not-found error from get_data. The refusal is correct behaviour, not a bug.

For cross-organisation queries, use a sister MCP:

• data360-mcp — World Bank's multi-source aggregator covering WDI, ILO, UN, and other official providers under one tool surface.

• worldbank-mcp — World Bank Open Data only.

For SDMX power-user queries against arbitrary registries (Eurostat, OECD, national NSOs), see sdmx-mcp — the generic SDMX protocol server discussed under Relationship to sdmx-mcp.

The scope boundary was sharpened in v1.1.1 after the v9 edge-test sample surfaced three prompts whose ground-truth codes (SI.POV.DDAY, NY.GDP.PCAP.PP.KD, per_allsp.cov_pop_tot) sit in the World Bank universe rather than the UNICEF Data Warehouse. Full write-up: internal/v1.1.0_design/ambiguity_forensic.md sections C-4 to C-6 (dev repo only).

Understanding UNICEF indicator codes

Why this matters — the MCP's semantic layer

UNICEF SDMX indicator codes look cryptic on first contact — PT_F_20-24_MRD_U18, NT_ANT_HAZ_NE2, TRGT_2030_IM_DTP3 — and most of an LLM's failure modes on this data start with picking the wrong code. The single biggest piece of added value this MCP offers over a raw SDMX endpoint is exposing the semantic structure inside those codes so the assistant can disambiguate without guessing. The implementation lives in src/unicefstats_mcp/differentiator.py (segment-level suffix meanings and base/variant explanation) and src/unicefstats_mcp/indicator_resolver.py (natural-language synonyms and known ambiguous tokens). v1.1.1 added a query-aware CURATED_PREFERRED dimension_hint that tells the resolver which code from a family to surface for a given phrasing (e.g. "target" → TRGT_*, "modelled" → *_MOD, "child" → _T total rather than a sex-disaggregated variant).

This section documents the conventions the MCP relies on so users and downstream agents can read codes directly, and so reviewers can audit the resolver's choices.

Anatomy of a UNICEF code

UNICEF DW codes are concatenations of dot-free, underscore-delimited segments, read left-to-right from broadest to most specific. Take a layered protection indicator from the catalog:

PT_F_20-24_MRD_U18
│  │ │     │   │
│  │ │     │   └── Marriage cutoff:        U18   = first union before age 18 (SDG 5.3.1)
│  │ │     └────── Indicator class:        MRD   = ever-married / in-union
│  │ └──────────── Age band (cohort):      20-24 = women 20-24 (retrospective denominator)
│  └────────────── Population restriction: F     = female-only indicator (women/girls)
└───────────────── Family prefix:          PT    = Child Protection

Read out loud: "Child Protection / female respondents / cohort aged 20-24 / ever-married / before age 18" — i.e. the share of women aged 20-24 who were first married before 18. Note that the F here is not a sex-disaggregation suffix appended to a sex-neutral parent; it is a population-restriction marker built into a family of indicators that only exist for female respondents (FGM, child marriage, anaemia in women, antenatal care). See "Population restrictions vs sex disaggregation" below.

The same left-to-right reading applies to the nutrition anthropometric pattern:

NT_ANT_HAZ_NE2
│  │   │   │
│  │   │   └── Threshold:    NE2 = below -2 SD (NE = "negative end", 2 = SD count)
│  │   └────── Metric:       HAZ = Height-for-Age Z-score
│  └────────── Sub-family:   ANT = Anthropometry
└───────────── Family:       NT  = Nutrition

NT_ANT_HAZ_NE2 is the stunting prevalence indicator for children under 5. To get the sex-stratified value, query the indicator with the SEX dimension filter (get_data(indicator='NT_ANT_HAZ_NE2', sex='F')), not by appending _F to the code — see the next subsection.

And the derived-metric pattern, which the v1.1.1 query-aware scoring is specifically tuned for:

TRGT_2030_IM_DTP3
│    │    │  │
│    │    │  └──── Antigen / dose: DTP3 = third dose of DTP
│    │    └─────── Family:         IM   = Immunization
│    └──────────── Target year:    2030 = SDG horizon
└───────────────── Derived metric: TRGT = country-set target value

The combinatorial reading — family → sub-family → metric → threshold/age band → disaggregation token — is the unwritten contract behind almost every code in the catalog. The tables below enumerate the parts.

Topic prefixes (families)

The first segment is the topical family. The top 10–12 prefixes in the catalog cover the vast majority of child-focused indicators:

The full mapping of natural-language phrases to these prefixes lives in indicator_resolver._SYNONYMS — e.g. "stunting" → NT_ANT_HAZ_NE2, "child marriage" → PT_F_20-24_MRD_U18, "BCG" → IM_BCG.

Methodology and provenance suffixes

A trailing segment often signals how the value was produced, not what was measured. These suffixes are central to the v1.1.1 query-aware scoring because users almost always want either "survey" or "modelled" but rarely both.

The MCP's differentiator.py:_SUFFIX_MEANINGS table is the canonical source. When two codes differ only in this trailing segment, explain_difference() reports them as base vs. variant (e.g. ECD_CHLD_LMPSL as base, _MERGE and _PRXY as derivation variants).

Population restrictions vs sex disaggregation

This distinction is easy to miss and routinely trips up downstream agents.

Population restriction (built into the code). Some UNICEF indicators only exist for one sex because the underlying measurement only applies to one sex: FGM prevalence, child-marriage cohorts, anaemia in women of reproductive age, antenatal care coverage. In those families, an F segment near the start of the code (PT_F_*, NT_ANE_WOM_*, MNCH_BIRTH18, antenatal-care codes) is a population-restriction marker that is part of the indicator's identity. There is no PT_M_20-24_MRD_U18 counterpart for child marriage; the indicator is defined on female respondents. Treat the F here as part of the indicator name, not as a disaggregation switch.

Sex disaggregation (SEX dimension filter, at query time). For indicators that are defined on both sexes (under-5 mortality, primary completion, stunting, literacy), sex is not encoded in the code. It is a separate SDMX dimension — the SEX dimension — with values F, M, _T (total). You select a slice at query time:

get_data(indicator='CME_MRY0T4', sex='F')   # under-5 mortality, girls
get_data(indicator='CME_MRY0T4', sex='M')   # under-5 mortality, boys
get_data(indicator='CME_MRY0T4', sex='_T')  # under-5 mortality, total

The differentiator.py:_SUFFIX_MEANINGS table lists F, M, _T, MF as sex-token meanings; those entries describe the values of the SEX dimension, not a suffix you append to an arbitrary code. Appending _F to a code that does not already carry it (e.g. inventing NT_ANT_HAZ_NE2_F) will not resolve — the catalog does not contain such codes.

Age / wealth / residence disaggregation tokens

Disaggregation tokens appear either embedded in the code (when they are part of the canonical definition, e.g. PT_F_20-24_MRD_U18) or applied at query time via get_data() filters. The tokens below are the same in both places.

Age bands

Age is encoded as LOW-HIGH (inclusive) in the embedded form, or as a bound token at the end of the code.

Wealth quintile (query-time filter)

Residence (query-time filter)

Anthropometric Z-score thresholds

These appear in the NT_ANT_* family and need their own table because the convention is non-obvious:

Derived metrics: TRGT_ / _ARR_ / _PRJ

Three special grammars flag values that are not observations but transformations of them. These are the cases where v1.1.1's query-aware CURATED_PREFERRED scoring matters most. Note that ARR is a middle-segment token (e.g. CME_ARR_U5MR, CME_ARR_SBR), not a trailing suffix.

v1.1.1 query-aware scoring (commit 7112e1d): the MCP only surfaces TRGT_* codes when your natural-language query mentions target, goal, objective, or a horizon year. A bare "DTP3 coverage in Brazil" question will resolve to the observation series; "Brazil's 2030 DTP3 target" will resolve to TRGT_2030_IM_DTP3. This is implemented as a hint field on CURATED_PREFERRED entries in differentiator.py, not as a hard exclusion — the target code is still discoverable via direct lookup, just demoted in resolver ranking unless the query signals intent.

The same query-aware demotion applies to _MOD (modelled) variants: queries mentioning survey, raw, or observed prefer the un-suffixed code; queries mentioning modelled, estimate, joint estimation prefer _MOD.

Education levels

In the ED_* family, ISCED levels are abbreviated L1 / L2 / L3:

Examples: ED_ANAR_L1 (adjusted net attendance rate, primary), ED_ANAR_L2 (lower-secondary), ED_CR_L1 (completion rate, primary). Where a code stops at L1 it is primary-only; where two levels are reported jointly the codes are listed separately rather than concatenated.

Scope caveat. The L<n> = ISCED <n> mapping holds only inside the ED_* family. Outside ED_*, L<n> may encode a non-ISCED level — for example PV_CHLD_MPI_L1 and PV_CHLD_MPI_L2 use L1 / L2 to mean severe and moderate multidimensional-poverty deprivation respectively. Some ED_* codes also use the two-digit form L01 / L02 to encode early-childhood or pre-primary levels that sit below ISCED 1. Always check the indicator's name before assuming L<n> means primary / lower-secondary / upper-secondary.

Where this is encoded in the MCP

The conventions documented above are not folklore — they are encoded in the MCP source and can be audited directly:

• src/unicefstats_mcp/differentiator.py is the canonical reference for segment-level meaning. The _SUFFIX_MEANINGS table maps every recognised trailing token to a human-readable gloss; explain_difference() walks two codes side-by-side and labels each diverging segment; the CURATED_PREFERRED table carries per-indicator dimension_hint strings that drive the v1.1.1 query-aware scoring.

• src/unicefstats_mcp/indicator_resolver.py maps natural-language phrases to canonical codes. _SYNONYMS covers the routine cases ("stunting", "under-5 mortality", "child marriage"); _AMBIGUOUS flags phrases that legitimately map to more than one code and require user disambiguation; _DISAMBIGUATION_TIPS carries the short hints surfaced in the assistant_guidance envelope field added in v1.1.0.

• unicef://glossary MCP resource — clients that load resources at session start get the disaggregation-code and indicator-prefix legend without having to parse this README.

When the resolver picks a code, the chain is: natural-language query → indicator_resolver lookup → if ambiguous, _AMBIGUOUS hit fires the v1.0.0 ambiguity_flag + candidate list → if a CURATED_PREFERRED entry matches, the dimension_hint re-ranks candidates → the chosen code is annotated by differentiator.py for the assistant_guidance field. Every step is inspectable in the source; nothing in this section is heuristic on the LLM side.

Demo

Step 1: Search for indicators

>>> search_indicators("stunting", limit=3)

{
  "query": "stunting",
  "total_matches": 11,
  "showing": 3,
  "results": [
    {"code": "FD_STUNTING", "name": "Moderate and severe stunting (Functional difficulties)"},
    {"code": "NT_ANT_HAZ_NE2", "name": "Height-for-age <-2 SD (stunting)"},
    {"code": "NT_ANT_HAZ_NE3", "name": "Height-for-age <-3 SD (severe stunting)"}
  ],
  "tip": "Use get_indicator_info('FD_STUNTING') for full details including available disaggregations."
}

Step 2: Get indicator metadata

>>> get_indicator_info("CME_MRY0T4")

{
  "code": "CME_MRY0T4",
  "name": "Under-five mortality rate",
  "description": "Probability of dying between birth and exactly 5 years of age, expressed per 1,000 live births",
  "dataflow": "GLOBAL_DATAFLOW",
  "sdmx_api": "https://sdmx.data.unicef.org/ws/public/sdmxapi/rest/data/UNICEF,GLOBAL_DATAFLOW,1.0/.CME_MRY0T4?format=csv",
  "disaggregation_filters": {
    "sex": ["_T (Total)", "M (Male)", "F (Female)"],
    "wealth_quintile": ["Q1 (Lowest)", "Q2", "Q3", "Q4", "Q5 (Highest)"],
    "residence": ["_T (Total)", "U (Urban)", "R (Rural)"]
  }
}

Step 3: Check temporal coverage

>>> get_temporal_coverage("CME_MRY0T4")

{
  "code": "CME_MRY0T4",
  "start_year": 1931,
  "end_year": 2024,
  "latest_year": 2024,
  "countries_with_data": 249,
  "note": "Not all countries have data for all years. Coverage varies by country."
}

Step 4: Fetch data

>>> get_data("CME_MRY0T4", ["BRA", "IND", "NGA"], start_year=2018, end_year=2023)

{
  "indicator": "CME_MRY0T4",
  "countries_requested": ["BRA", "IND", "NGA"],
  "total_rows_available": 18,
  "rows_returned": 18,
  "rows_truncated": false,
  "format": "compact",
  "summary": {
    "value_range": {"min": 14.42, "max": 117.56, "mean": 54.78},
    "year_range": {"earliest": 2018, "latest": 2023},
    "countries_in_result": 3
  },
  "data": [
    {"iso3": "BRA", "country": "Brazil",  "period": 2018, "indicator": "CME_MRY0T4", "value": 15.22},
    {"iso3": "BRA", "country": "Brazil",  "period": 2019, "indicator": "CME_MRY0T4", "value": 15.03},
    {"iso3": "BRA", "country": "Brazil",  "period": 2020, "indicator": "CME_MRY0T4", "value": 14.87},
    {"iso3": "BRA", "country": "Brazil",  "period": 2021, "indicator": "CME_MRY0T4", "value": 14.72},
    {"iso3": "BRA", "country": "Brazil",  "period": 2022, "indicator": "CME_MRY0T4", "value": 14.59},
    {"iso3": "BRA", "country": "Brazil",  "period": 2023, "indicator": "CME_MRY0T4", "value": 14.42},
    {"iso3": "IND", "country": "India",   "period": 2018, "indicator": "CME_MRY0T4", "value": 36.87},
    {"iso3": "IND", "country": "India",   "period": 2019, "indicator": "CME_MRY0T4", "value": 34.86},
    {"iso3": "IND", "country": "India",   "period": 2020, "indicator": "CME_MRY0T4", "value": 32.98},
    {"iso3": "IND", "country": "India",   "period": 2021, "indicator": "CME_MRY0T4", "value": 31.19},
    {"iso3": "IND", "country": "India",   "period": 2022, "indicator": "CME_MRY0T4", "value": 29.53},
    {"iso3": "IND", "country": "India",   "period": 2023, "indicator": "CME_MRY0T4", "value": 27.99},
    {"iso3": "NGA", "country": "Nigeria", "period": 2018, "indicator": "CME_MRY0T4", "value": 117.19},
    {"iso3": "NGA", "country": "Nigeria", "period": 2019, "indicator": "CME_MRY0T4", "value": 117.37},
    {"iso3": "NGA", "country": "Nigeria", "period": 2020, "indicator": "CME_MRY0T4", "value": 117.42},
    {"iso3": "NGA", "country": "Nigeria", "period": 2021, "indicator": "CME_MRY0T4", "value": 117.56},
    {"iso3": "NGA", "country": "Nigeria", "period": 2022, "indicator": "CME_MRY0T4", "value": 117.46},
    {"iso3": "NGA", "country": "Nigeria", "period": 2023, "indicator": "CME_MRY0T4", "value": 116.82}
  ]
}

Key insights an AI assistant would extract from this:

• Brazil: 14.4 per 1,000 — steadily declining, on track for SDG 3.2 target (≤25)

• India: 28.0 per 1,000 — rapid improvement (37→28 in 5 years), recently crossed SDG target

• Nigeria: 117 per 1,000 — essentially flat, 4.7× the SDG target, highest burden

Step 5: Get code template to continue in a script

>>> get_api_reference("r", "unicefData")

{
  "language": "r",
  "install": "install.packages(\"unicefdata\")",
  "import": "library(unicefdata)",
  "function": "unicefData",
  "signature": "unicefData(\n    indicator = NULL,        # character — indicator code(s)\n    countries = NULL,         # character vector — ISO3 codes, NULL = all\n    year = NULL,              # numeric, character (\"2015:2023\"), or vector\n    sex = \"_T\",               # character — \"_T\", \"M\", \"F\"\n    totals = FALSE,           # logical — only return aggregate totals\n    tidy = TRUE,              # logical — standardize column names\n    country_names = TRUE,     # logical — add country name column\n    format = \"long\",          # character — \"long\", \"wide\", \"wide_indicators\"\n    latest = FALSE,           # logical — most recent value per country\n    circa = FALSE,            # logical — closest available year\n    add_metadata = NULL,      # character vector — e.g. c('region', 'income_group')\n    dropna = FALSE,           # logical — drop rows with missing values\n    simplify = FALSE,         # logical — minimal columns\n    mrv = NULL,               # integer — most recent N values per country\n    raw = FALSE,              # logical — all disaggregations, no filtering\n)",
  "returns": "tibble with columns: indicator_code, iso3, country, period, value, sex, age, wealth_quintile, residence, ...",
  "examples": [
    {"description": "Under-5 mortality for Brazil, India, Nigeria (2015–2023)", "code": "df <- unicefData(\"CME_MRY0T4\", countries = c(\"BRA\", \"IND\", \"NGA\"), year = \"2015:2023\")"},
    {"description": "Latest stunting data for all countries", "code": "df <- unicefData(\"NT_ANT_HAZ_NE2\", latest = TRUE)"},
    {"description": "Wide format with region metadata", "code": "df <- unicefData(\"CME_MRY0T4\", format = \"wide\", add_metadata = c(\"region\", \"income_group\"))"}
  ]
}

This lets the AI generate correct R/Python/Stata code using the exact parameter names and syntax — no guessing from training data.

get_data parameters

Response features

• summary: Value range (min/max/mean), year range, country count

• disaggregations_in_data: Which dimensions have non-trivial variation

• total_rows_available vs rows_returned: Pagination metadata

• tip: Contextual guidance for next steps or narrowing results

Prompts

compare_indicators

Pre-built analysis workflow: fetches indicator metadata and data, then produces a structured comparison.

compare_indicators(indicator="CME_MRY0T4", countries="BRA,IND,NGA", start_year="2015", end_year="2023")

write_unicefdata_code

Generate runnable Python, R, or Stata code using the unicefdata package. The AI will call get_api_reference() to get the exact function signatures, then write code matching the user's task.

write_unicefdata_code(
    task="Compare under-5 mortality for Brazil and India, 2015-2023, then plot the trends",
    language="r"
)

This bridges the gap between conversational exploration (via MCP tools) and reproducible analysis scripts (via unicefdata packages).

Benchmark Results

We benchmarked the MCP against a bare LLM (Claude Sonnet 4, no tools) using the EQA metric from Azevedo (2025). 300 queries across 10 indicators, 20 countries, 2 prompt types, and 2 hallucination test categories.

Current canonical numbers (v0.7.3 + fixes, May 2026)

The numbers below are the current canonical scoreboard. They reflect (a) four engineering fixes in the v0.7.3 cycle (see CHANGELOG) and (b) a scoring correction from the v1.4 extractor that respects refusal language. Both samples (mcp060: 40 countries; mcp073: disjoint 20-country validation sample) score under the same v1.4 rules.

v0.7.3 + fixes is the first version where MCP demonstrably makes the model safer than the no-tools baseline on absent-data queries. Through v0.7.2, hall_b ≥ hall_a — the safety layer was reducing magnitude but not direction. The four fixes that flipped the property:

1. server.py:_seed_data_frontier_cache — monotonic max instead of unconditional overwrite (a probe-induced regression of the cached frontier was telling the LLM that current data was unavailable).

2. server.py:get_data exception handler — unicefdata cascade exhaustion reclassified as no_data (not error), so the LLM treats it as authoritative absence rather than tool failure.

3. benchmark_eqa_batch.py — persist tool result_str on state.tool_calls so the v1.4 extractor can see refusals.

4. benchmark_eqa_batch.py — refusal-respect parity with the sync runner.

Canonical scoreboard: see the [Unreleased] §Fixed block at the top of CHANGELOG.md for the full cross-version table, including the four post-fix corrections and the mcp073 second-sample validation. Full per-run write-ups live in internal/v0_7_3_validation.md and internal/v0_7_3_second_sample_validation.md in the dev repo (jpazvd/unicefstats-mcp-dev, dev-only and not synced to this public mirror).

Historical: v0.3.0 (n=600, 2025) and v0.7.2 (n=500 same-day, 2026-05-08)

The original v0.3.0 benchmark reported POS EQA 0.147 → 0.990 (6.7×) and T2 hallucination 11% → 37%. Two corrections since:

• The "37%" headline was substantially a v1.3 extractor scoring artefact: the extractor counted any numeric value mentioned in the response as a "claim," including values quoted from no_data tool results inside an explicit refusal. The v1.4 extractor (_detect_refusal) reclassifies those as appropriate refusals. We did not rescore the v0.3.0 parquets under v1.4 (they're archived); the v1.4-equivalent of "37%" is unknown but substantially lower.

• The v0.7.2 reproduction (n=500, 2026-05-08) under v1.3 scoring reported POS EQA 0.897 and combined T1+T2 hallucination 13%. Under v1.4 scoring the same parquets report POS EQA 0.793 and hall_b 3.75%.

The accuracy headline (~7× lift) has held up across every rescoring. The hallucination headline required both a scoring fix (v1.4 extractor) and a server fix (v0.7.3 cache + cascade fixes) before MCP actually reduced hallucination below the no-tools baseline.

EQA decomposition (baseline_latest prompt)

Key findings

1. All 10 indicators at EQA >= 0.95 with MCP, replicated across 40 countries (R1 + R2 with zero overlap). 7 of 10 achieve perfect EQA = 1.000.

2. Year accuracy is the bare LLM's biggest weakness (YA = 0.24). It cites 2021-2022 as "latest" when IGME 2024 estimates exist. The MCP queries the API and returns the actual latest year.

3. The direct prompt shows larger MCP gain (+0.722 vs +0.613) because it eliminates YA and isolates pure retrieval accuracy.

4. T2 hallucination (~37%) is inflated by ground truth misclassification: the SDMX API has IGME mortality data for micro-states that the ground truth pipeline missed. After correction: MCP ~10%, LLM alone ~5%. The remaining hallucination is driven by the confidence effect — Claude overrides tool errors when it has strong domain priors.

5. The confidence effect: When the MCP tool returns "no data" but the LLM has strong domain priors (e.g., child mortality for well-known countries), it overrides the tool and fabricates anyway. This is a fundamental LLM behavior, not MCP-specific.

3-way comparison (vs sdmx-mcp)

sdmx-mcp's raw SDMX-JSON output is hard for LLMs to parse (VA ≈ 0.11), but its anti-hallucination guardrails are highly effective (0% fabrication). See Relationship to sdmx-mcp for details.

Full analysis, per-indicator decomposition, and methodology: examples/RESULTS.md

Benchmark data (parquet with full LLM responses): examples/results/

Benchmark design rationale: examples/DESIGN_ISSUES.md

Reproducing the benchmark

# Build ground truth from UNICEF SDMX API
python examples/00_build_ground_truth.py

# Run 200-query benchmark (requires ANTHROPIC_API_KEY, ~$6)
python examples/benchmark_eqa.py

# Add 100 direct-prompt queries to existing run (~$3)
python examples/01_run_direct_supplement.py

Citation

This benchmark uses the EQA metric from:

Azevedo, J.P. (2025). "AI Reliability for Official Statistics: Benchmarking Large Language Models with the UNICEF Data Warehouse." UNICEF Chief Statistician Office. github.com/jpazvd/unicef-sdg-llm-benchmark-dev

Deployment

Local (stdio)

unicefstats-mcp

Remote (SSE)

unicefstats-mcp --transport sse --port 8000

Docker

docker build -t unicefstats-mcp .
docker run -p 8000:8000 unicefstats-mcp

Development

pip install -e ".[dev]"
pytest tests/ -v
ruff check src/ tests/
mypy src/unicefstats_mcp/

Contributing

Contributions are welcome.

Ways to contribute

• Bug reports: Open an issue with steps to reproduce

• Feature requests: Suggest new tools, indicators, or output formats via issues

• Code: Fork, branch, submit a PR — see development setup below

• Benchmark: Run the EQA benchmark on different models and share results

• Documentation: Improve examples, fix typos, add use cases

Development setup

git clone https://github.com/jpazvd/unicefstats-mcp.git
cd unicefstats-mcp
pip install -e ".[dev,benchmark]"
pytest tests/ -v
ruff check src/ tests/
mypy src/unicefstats_mcp/

Pull request guidelines

1. One concern per PR — keep changes focused and reviewable

2. Include tests for new tools or bug fixes

3. Run the linter (ruff check) and type checker (mypy) before submitting

4. Update the README if you change tool signatures or add new features

5. Do not commit API keys or benchmark result parquets larger than 500KB

Priority areas

See the audit findings for known issues. High-impact areas:

• MNCH dataflow bug: MNCH_CSEC and MNCH_BIRTH18 return 0 EQA due to a dataflow resolution issue in the unicefdata package

• T2 hallucination reduction: Further reduce fabrication when API returns no results (currently ~10%; see Limitations)

Limitations and Hallucination Risks

Data limitations

• Coverage is uneven across indicators, countries, and years. Survey-based indicators (nutrition, education, protection) have 3-5 year gaps between data points by design.

• Mortality indicators (CME_*) are modeled estimates from the UN Inter-agency Group (IGME), with uncertainty intervals not surfaced in compact output.

• Not all indicators support all disaggregation dimensions; get_indicator_info() lists what's available per indicator.

• get_data() caps at 500 rows per call.

Hallucination risks

Benchmark testing across multiple country samples (v0.7.3 + fixes, May 2026):

v0.7.3 + fixes is the first release where MCP makes the model safer than the no-tools baseline. Through v0.7.2 (v1.4 scoring), hall_b ≥ hall_a — the safety layer was reducing magnitude relative to a no-safety-layer baseline but never enough to put MCP under the no-tools floor.

What changed:

1. Cache contamination in _seed_data_frontier_cache — a probe routine was overwriting the cached max-year for an indicator with whatever year happened to come back, sometimes lower than the cached value. The LLM was being told frontiers were 2018 for indicators whose real frontier was 2023, then answering about 2023 from parametric memory. Fix: monotonic max() only.

2. unicefdata cascade as no_data, not error — the underlying wrapper's fallback-dataflow exhaustion was leaking as a tool exception, which the LLM read as "tool failed, fall back to my own knowledge." Now classified as a clean no_data signal that the safety layer converts into "do not estimate." Filed unicefdata-dev#74 for an upstream fix.

3. v1.4 extractor (_detect_refusal) — corrected the long-standing scoring artefact in which any numeric value in a response was counted as a "claim," including values the LLM was quoting from no_data tool results inside an explicit refusal. This dropped hall_b in the v0.7.3 PRE-FIX rescoring from ~37% to 2.25% — most of the historical "MCP-makes-it-worse" finding was extractor, not behaviour.

4. Batch runner parity — result_str persistence and refusal-respect in the async batch runner.

This finding still leaves room for the broader tool-augmented LLM literature on the structural cost of giving models an answer-producing pathway:

• The Reasoning Trap: How Enhancing LLM Reasoning Amplifies Tool Hallucination (ICLR 2025) — shows the relationship is causal: as models get better at tool use, tool hallucination rises proportionally with capability.

• Reducing Tool Hallucination via Reliability Alignment (Cao et al., 2024, arXiv:2412.04141) — formalises the failure as tool-selection errors (wrong tool, failed refusal) and tool-usage errors (fabricated parameters).

• ReDeEP: Detecting Hallucination in Retrieval-Augmented Generation via Mechanistic Interpretability (Sun et al., 2024) — shows mechanistically that an LLM's parametric knowledge can override retrieved context inside the residual stream.

These results show the structural tendency is real; the v0.7.3 + fixes result shows it can be reversed for a specific data domain through (a) safety-layer architecture, (b) server-side discipline (no stale frontier cache, no leaking of no-data as tool errors), and (c) honest scoring.

The takeaway for users:

1. Load the unicef://system-prompt and unicef://context resources at session start (handles forward-of-frontier fabrication).

2. Treat MCP results as best-effort retrieval, not infallible truth — verify load-bearing values against the UNICEF Data Warehouse before citing.

3. Prefer queries with explicit years ("under-five mortality in Nigeria in 2023") over open-ended ones ("the latest under-five mortality in Nigeria") — the former triggers refusal more reliably when data is absent.

4. The 1.00% / 2.25% residuals were measured on Sonnet 4 only. Cross-model generalisation is the next benchmark.

Full benchmark methodology: examples/RESULTS.md

Provenance and Ownership

All data served by this MCP originates from the UNICEF Data Warehouse, accessed live via the public SDMX REST API. No observation data is stored or cached — every get_data() call results in a live SDMX request. The indicator and country registries are cached in memory at first access for performance; these are catalogue metadata, not statistical values. The MCP reformats output for LLM consumption but does not alter values.

All releases are published from GitHub Actions using PyPI Trusted Publishing (OIDC). No long-lived API tokens exist. Release provenance is verifiable via PyPI attestations.

For full details on data origin, ownership, distribution pipeline, and interpretation caveats, see PROVENANCE.md.

How to Verify This MCP

License

MIT