Tools

9 tools for browsing and querying the World Bank Open Data API:

worldbank_list_topics

List all 21 World Bank thematic topic categories.

• No input required — returns the complete fixed taxonomy
• Topics include Agriculture, Economy & Growth, Education, Energy & Mining, Environment, Financial Sector, Health, Infrastructure, Poverty, Private Sector, Public Sector, Science & Technology, Social Development, Social Protection & Labor, Trade, Urban Development, and more
• Returns topic IDs used to filter worldbank_search_indicators

worldbank_list_sources

List the 70+ World Bank data sources (datasets).

• Paginated with configurable page size (up to 100 per page)
• Each source includes ID, name, short code, last-updated date, and data/metadata availability status
• Source IDs used to filter worldbank_search_indicators by dataset origin (e.g. "2" for World Development Indicators, "6" for IDS)

worldbank_list_countries

List countries and regional aggregates with metadata.

• Returns ISO codes, World Bank region, income level, capital city, and coordinates
• Filterable by region code (EAS, ECS, LCN, MEA, NAC, SAS, SSF) and income level (LIC, LMC, UMC, HIC)
• By default returns individual countries only; set include_aggregates=true to include regional, income-group, and world aggregate entities
• Paginated with up to 300 entries per page

worldbank_get_country

Fetch full metadata for a single country or aggregate entity.

• Accepts ISO2 (US), ISO3 (USA), or World Bank aggregate codes (EAS, HIC, WLD)
• Returns region, income level, lending type, capital city, and coordinates
• Structured error with recovery hint when code is not found

worldbank_search_indicators

Search the 29,500+ World Bank indicator catalog.

• Keyword search, topic filter, source filter — at least one required
• The upstream searchterm parameter does not filter, so keyword matching runs locally over the full catalog (or the full selected topic/source): every term must appear in the indicator ID, name, or description, in any word order, with punctuation ignored
• Exact ID or name matches rank first, then whole-phrase matches, then other ID/name matches, then description-only matches — so pasting an indicator name (GDP (current US$)) or ID (NY.GDP.MKTP.CD) returns it as the top hit
• Returns indicator IDs, names, source dataset, and thematic topics
• One row per indicator ID — the 43 indicators published under both a live source and an archived copy collapse to the live row, and worldbank_get_indicator resolves to the same one
• Indicator IDs (e.g. NY.GDP.PCAP.CD, SP.POP.TOTL) feed directly into worldbank_get_data
• Echoes the filters it applied (query, topic_id, source_id) alongside the results
• Paginated with up to 100 results per page

worldbank_get_indicator

Fetch complete metadata for a known indicator ID.

• Returns full description, unit of measurement, source dataset, source organization, and thematic topics
• Structured error with recovery hint when ID is not found

worldbank_get_data

Query indicator values for countries across time. The primary data-access tool.

• Single country, array of countries, regional codes (EAS, LCN, …), income codes (HIC, LMC, …), world code (WLD), or "all" for every entry. At least one code is required — an empty value is rejected rather than read as "all"
• Time filtering: date_range for a period or range — annual (2020, 2010:2023), quarterly (2020Q1:2021Q4), or monthly (2020M01:2020M06) — or mrv for the N most recent available values (1–100). Mutually exclusive.
• A date_range the API can't apply is enforced locally over the whole series, so observations outside the requested window are never returned as matches
• Returns observations with null values when data is not available for a country×year cell — common for sparse series
• Includes nullCount per page to surface data sparsity
• Output grouped by country for readability; isAggregate flag distinguishes all 78 regional, income-group, and lending-group aggregates from individual countries
• Echoes the parameters it sent upstream — indicator, normalized country codes, date range or mrv, page and page size
• Paginated with up to 1000 entries per page

worldbank_get_poverty

Poverty and inequality estimates from the World Bank Poverty and Inequality Platform (PIP) — a separate dataset from the WDI series the other tools read.

• Poverty headcount ratio, poverty gap, poverty severity, and the Watts index at any poverty line, not only the published thresholds. Omit poverty_line to use the international line of the PIP release currently served, which is echoed back on every row
• The same row carries the inequality and distribution block: Gini coefficient, mean log deviation, Wolfson polarization, and the ten decile income/consumption shares. One tool, because PIP returns poverty and inequality together and the poverty line does not affect the distribution
• Every row states how it was produced. estimationType: "survey" rows carry the full inequality block; "interpolation", "extrapolation", and "CMD estimation" rows are gap-filled estimates for years no survey covers, and their gini, mld, polarization, and decileShares are null — a gap in the source data, not an error. surveyYear, surveyAcronym, and isInterpolated sit alongside them
• fill_gaps defaults to true, so years the surveys don't reach come back as PIP's own estimate rather than as an empty result or a hole in the series. Survey rows win wherever one exists, so a single-year batch can carry real distributions for some economies and labelled estimates for others, and a full-history query returns a row per year with the real distribution on the survey years. Set fill_gaps=false for survey-derived rows only
• Welfare is measured in PPP dollars per person per day. welfare_type pins results to income or consumption surveys, which are not directly comparable; thirty-five economies publish both and return a row for each
• Ten economies publish an urban/rural split and return an extra row per year for it — China alone reports all three levels, the rest pair national with either urban or rural. reporting_level narrows to one, and every row states its own
• year accepts a four-digit year, "all" for the full history, or "MRV" for the most recent available. Coverage starts in 1963
• Individual economies only, by ISO3 code — PIP does not serve regional, income-group, or world aggregate codes through this endpoint, and a structured error says so
• Paginated locally with up to 1000 entries per page, since PIP itself has no pagination

worldbank_search_projects

The World Bank lending portfolio — the individual loans, credits, and grants the Bank finances — rather than the statistics the other tools read. A third upstream API, on its own host, with its own envelope and conventions.

• Free-text query across project names, abstracts, and objectives, combined with exact filters on countries, region, status, and a board-approval date window. Every filter is an AND, and every multi-value filter is an OR within itself
• Returns the project ID, name, borrowing country and region, status, board approval and closing dates, total commitment in USD, financing windows (IBRD, IDA, Grants), major sectors, and a link to the project page
• Countries are ISO2 here (BR, IN, ZA) — the one place this server departs from the ISO3 codes its other tools take, because the Projects API keys on ISO2. Multi-country operations carry a World Bank regional code instead (3A, 4E), two characters like the rest. An ISO3 code is rejected at the schema with a message naming the difference, rather than being passed upstream where it would come back as a silent zero-hit. worldbank_get_country reports a country's iso2 field for either form
• status is a closed set (Active, Closed, Dropped, Pipeline) and region is the nine World Bank operational regions, both enforced by the schema — a typo in either can't reach upstream and read back as "no results"
• An empty result says which filter emptied it. When a country filter was in force, the search re-runs it on its own: either no project carries those codes at all, or they match as a set and the remaining filters are what narrowed the result to nothing, and the notice names them
• include_abstract is off by default. Abstracts run to a median of roughly 1,200 characters and are published for about half the portfolio, so a full page of them roughly doubles the response — turn it on once the result set is small enough to read
• Offset-paginated with up to 1000 entries per page, which is also the most the API returns for one request

Resources

Both resources return a structured not-found error with a recovery hint for an unknown ID or code. An upstream outage, timeout, or 5xx keeps its own classification, so a transient failure is distinguishable from a bad identifier.

Features

Built on @cyanheads/mcp-ts-core:

• Declarative tool definitions — single file per tool, framework handles registration and validation
• Unified error handling across all tools
• Pluggable auth (none, jwt, oauth)
• Swappable storage backends: in-memory, filesystem, Supabase, Cloudflare KV/R2/D1
• Structured logging with optional OpenTelemetry tracing
• Runs locally (stdio/HTTP) or on Cloudflare Workers from the same codebase

World Bank-specific:

• Full World Bank Open Data API v2 coverage — topics, sources, countries, indicators, and observations
• 60+ years of development data across 29,500+ indicators for 200+ countries and regional aggregates
• Client-side topic/source + keyword compound filtering (works around upstream API limitation)
• Null-value transparency — null observations and nullCount surfaced rather than silently dropped
• isAggregate flag on every country/data row to distinguish individual countries from aggregate entities

Agent-friendly output:

• Tool cross-references woven into descriptions — e.g. worldbank_search_indicators names worldbank_list_topics for topic IDs, worldbank_get_data names worldbank_search_indicators for indicator discovery
• Structured error contracts with typed reason codes and actionable recovery hints on every tool
• Consistent pagination metadata (page, pages, total) across all list/search/data tools

Getting started

Self-Hosted / Local

Add the following to your MCP client configuration file.

{
  "mcpServers": {
    "worldbank-mcp-server": {
      "type": "stdio",
      "command": "bunx",
      "args": ["@cyanheads/worldbank-mcp-server@latest"],
      "env": {
        "MCP_TRANSPORT_TYPE": "stdio",
        "MCP_LOG_LEVEL": "info"
      }
    }
  }
}

Or with npx (no Bun required):

{
  "mcpServers": {
    "worldbank-mcp-server": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@cyanheads/worldbank-mcp-server@latest"],
      "env": {
        "MCP_TRANSPORT_TYPE": "stdio",
        "MCP_LOG_LEVEL": "info"
      }
    }
  }
}

For Streamable HTTP, set the transport and start the server:

MCP_TRANSPORT_TYPE=http MCP_HTTP_PORT=3010 bun run start:http
# Server listens at http://localhost:3010/mcp

Prerequisites

• Bun v1.3.0 or higher (or Node.js ≥24.0.0).
• No API key required — the World Bank Open Data API is public and unauthenticated.

Installation

1. Clone the repository:

git clone https://github.com/cyanheads/worldbank-mcp-server.git

2. Navigate into the directory:

cd worldbank-mcp-server

3. Install dependencies:

bun install

Configuration

All configuration is validated at startup via Zod schemas in src/config/server-config.ts. Environment variables:

Running the server

Local development

• Build and run the production version:

  # One-time build
  bun run rebuild
  
  # Run the built server
  bun run start:http
  # or
  bun run start:stdio
  

• Run checks and tests:

  bun run devcheck  # Lints, formats, type-checks, and more
  bun run test      # Runs the test suite
  

Project structure

Development guide

See CLAUDE.md for development guidelines and architectural rules. The short version:

• Handlers throw, framework catches — no try/catch in tool logic
• Use ctx.log for logging, ctx.state for storage
• Register new tools and resources in the createApp() arrays

Contributing

Issues and pull requests are welcome. Run checks and tests before submitting:

bun run devcheck
bun run test

License

This project is licensed under the Apache 2.0 License. See the LICENSE file for details.