Data360 MCP Server · World Bank

Overview

The Data360 MCP Server exposes the World Bank's Data360 Platform to any MCP-capable client. LLMs and agents need a structured way to search indicators, check metadata and disaggregation, fetch time-series data, and produce charts—without embedding Data360-specific logic inside each client. This server fills that role: it implements the Model Context Protocol so clients can discover and call a fixed set of tools, read contextual resources (system prompts, codelists, schemas), and optionally render tool results in interactive app UIs.

Every value returned by the server comes from the Data360 platform. The server is stateless with respect to user sessions—it acts as a bridge between MCP clients and the Data360 HTTP API.

Who this is for

Click an audience to see typical setup, tools, and resources.

How they use it
How they use it
How they use it

What your agent can do

An AI agent connected over MCP can query official World Bank series from Data360: search the catalog, pull time series, read indicator metadata, and request charts. Every value comes from the Data360 platform, mitigating the risk of hallucination.

Countries

Search indicators and confirm coverage for the countries named in the prompt.
Time series

Retrieve values by country, year, and breakdowns such as sex, age, or urban/rural.
Metadata

Surface source, definition, methodology, and known limitations for an indicator.
Charts

Build line, bar, or other charts and return a URL the MCP client can display.

Example questions

Sample prompts for a connected agent in Cursor, VS Code, or another MCP client. The agent resolves countries, selects indicators, and checks coverage before answering.

Click a question to see the likely MCP tool sequence an agent would run.

Likely tool sequence
Likely tool sequence
Likely tool sequence
Likely tool sequence

Charts and visualizations

For charts, the agent calls data360_get_viz_spec (single indicator) or data360_get_multi_indicator_viz_spec (2–4 indicators). The server loads series from Data360 and returns Vega-Lite JSON plus a chart URL. For tables or numbers in text, use data360_get_data or the aggregation tools listed under Technical reference—they are separate calls.

Chart flow: Search for the indicator, call data360_get_disaggregation for valid years and breakdowns, then request the chart spec.

Line and area — trends over years or periods
Bar — discrete periods or categories when appropriate
Scatter / point — relationships when the data supports it

Charts use Vega-Lite. Your MCP client opens the URL the server returns.

Data sources and databases

The server's only persistent data source is the World Bank Data360 HTTP API. All indicator search, metadata, disaggregation, and time-series data are fetched from Data360 endpoints. There is no local database—the server is stateless aside from in-memory caches (e.g. codelists). All Data360 databases are supported. Use data360_list_indicators to discover indicators in any database, or load data360://databases for a reference list.

Agent-safe design

The server is designed to reduce common LLM data errors through composable tools, coverage validation, and curated resources.

Coverage checks

Use required_country in search and data360_get_disaggregation before fetching data so the agent only requests series that exist.
Composable tools

Small, focused tools validate codes, filters, and dimensions before returning data—search first, then refine.
Bundled resources

System prompt, codelists, and usage notes under data360:// keep client logic in sync with server conventions.
Payload trimming

Use select_fields on metadata calls to limit response size; data payloads are pre-filtered to essential columns for LLM consumption.
Aggregation helpers

data360_summarize_data, data360_rank_countries, and data360_compare_countries return compact summaries instead of raw series when the question is about trends, rankings, or cross-country comparison.

Server capabilities

Search, metadata, series retrieval, code lists, and bundled prompt resources exposed to MCP clients.

Indicator search

Full-text search with metadata; optional country filter to limit results to series that actually exist.
Metadata

Methodology, definitions, limitations, and statistical concepts for each indicator.
Time series

Historical values with filters for country, period, sex, age, urbanization, and other dimensions.
Bundled resources

System prompt, codelists, and usage notes exposed as MCP resources under data360://.
Composable tools

Small, focused tools that validate coverage, codes, and filters before returning data.

MCP tools

Typical order: search (or search datasets), optional disaggregation, then get_data, aggregation tools, or viz specs. Do not guess database_id or indicator_id—obtain them from search results first.

Discovery and data

Tool	Description
`data360_search_indicators`	Search indicators with enriched metadata. Use `required_country` for coverage checks. Supports single `query`, multi-topic `queries`, or scoped `query_groups`. Returns `covers_country`, `latest_data`, `dimensions`.
`data360_search_datasets`	Search Data360 dataset catalogs (e.g. WDI, Findex) when the user asks about sources or databases by name.
`data360_get_metadata`	Indicator metadata and optional disaggregation preview; use `select_fields` to limit payload.
`data360_get_disaggregation`	Available filter values (countries, years, dimensions) per indicator. Call before `get_data` or chart tools.
`data360_get_data`	Fetch observations with country, year, and dimension filters (SEX, AGE, URBANISATION, etc.). Paginated.
`data360_find_codelist_value`	Resolve names to codes (e.g. "Kenya" → `KEN`, "female" → `F`).
`data360_expand_country_group`	Expand a REF_AREA group (region, income, lending) into constituent country codes (e.g. `SAS`, `LIC`).
`data360_list_indicators`	List all indicator IDs for a database.
`data360_get_data_api_url`	Low-level: build a direct Data360 data API URL.

Analysis and visualization

Tool	Description
`data360_summarize_data`	Summary statistics grouped by dimensions—use for trend or change-over-time questions without pulling full series.
`data360_rank_countries`	Rank countries by indicator value for a year (top/bottom performers, leaderboards).
`data360_compare_countries`	Compare 2–8 countries on one indicator (snapshot or aligned time series).
`data360_get_viz_spec`	Vega-Lite spec and chart URL for one indicator. Call `get_disaggregation` first. Fetches its own data; does not consume `get_data` output.
`data360_get_multi_indicator_viz_spec`	Chart comparing 2–4 indicators (scatter, dual-axis line, etc.).
`data360_get_supported_chart_types`	List supported chart types and data requirements.

Recommended agent workflow

Tables / numbers in text:
1. data360_search_indicators(query, required_country="Kenya")
2. data360_get_disaggregation(database_id, indicator_id)  — years and dimensions
3. data360_get_data(database_id, indicator_id, country_code="KEN", ...)

Rankings / comparisons (prefer over raw get_data when possible):
1. data360_search_indicators(...)
2. data360_get_disaggregation(...)
3. data360_rank_countries(...) or data360_compare_countries(...)

Charts (single indicator):
1. data360_search_indicators(...)
2. data360_get_disaggregation(...)
3. data360_get_viz_spec(database_id, indicator_id, country_code, ...)

Charts (multiple indicators):
1. data360_search_indicators(...) for each topic
2. data360_get_multi_indicator_viz_spec(indicator_ids=[...], ...)

MCP resources

Read-only URIs for prompts and reference data. Most integrations load data360://system-prompt and data360://context in the system message; larger references can be fetched on demand.

Resource	Description
`data360://system-prompt`	Required. Workflow and tool-use guidance for the system message.
`data360://context`	Recommended. Runtime context (current date and year) for time-aware queries.
`data360://agent-recipe`	Host integration recipe for LangGraph / `data360-mcp-agent` (resource + prompt composition).
`data360://k360-narrative-style`	Optional markdown response contract for staged narrative renderers.
`data360://databases`	Available databases (live mapping from Data360)
`data360://codelists`	Codelist reference (REF_AREA, SEX, AGE, UNIT_MEASURE, …)
`data360://metadata-fields`	Field mapping for smart question routing
`data360://data-filters`	Available filters and usage guidance
`data360://data-schema`	Standard columns and visualization guidance
`data360://search-usage`	Search examples and best practices

MCP prompts

Parameterized playbooks registered with prompts/list and prompts/get. Hosts fetch a prompt by name and prepend the returned text to the conversation for a specific user turn.

Prompt	When to use
`indicator_search`	Pick one indicator among search hits
`indicator_details`	Methodology, definition, or limitation questions
`country_data`	One country + theme → data and optional chart
`gate_classifier`	Decide in/out of scope before the tool loop
`thematic_to_data`	Rewrite a broad development question into a data task
`k360_research_compiler`	Build a JSON content packet from a tool trace
`k360_narrative`	Convert a content packet into narrative markdown

See data360://agent-recipe for recommended composition order with LangGraph and data360-mcp-agent.

How the server works

MCP clients connect to the Data360 MCP server over streamable HTTP. The server does not store user sessions; each request is independent. It calls the Data360 API for search, metadata, disaggregation, and data, and optionally an external Charts API for persisting Vega-Lite specs.

  MCP clients (Cursor, LangGraph, custom apps)
              │
              ▼ streamable HTTP /mcp
  ┌───────────────────────────────┐
  │     Data360 MCP Server        │
  │  FastMCP · Tools · Resources  │
  │  · Prompts · MCP Apps         │
  └───────────┬───────────────────┘
              │
     ┌────────┴────────┐
     ▼                 ▼
 Data360 HTTP API   Charts API (optional)
 search · metadata  Vega-Lite spec storage
 disagg · data

Layer	Technology	Purpose
MCP	FastMCP, MCP SDK	Tools, resources, prompts, apps
Transport	Streamable HTTP (stateless)	Production and local development
API client	httpx (async)	Data360 search, metadata, data, codelists
Visualization	Draco, Altair, Vega-Lite	Chart generation from time-series data

For the full architecture document, see architecture-data360-mcp.md.

MCP Apps

MCP Apps allow hosts to render tool results in interactive iframes instead of plain JSON. Tools attach a resourceUri in their metadata so the host can open the corresponding app and pass the tool result.

App	URI	Tool
Chart view	`ui://data360/chart-view.html`	`data360_get_viz_spec`
Search results	`ui://data360/search-results.html`	`data360_search_indicators`
QR view	`ui://data360/qr-view.html`	(example)

See MCP Apps implementation notes for alignment with the official MCP ext-apps extension.

Client libraries and examples

The repository includes packages and examples for common integration patterns.

data360-mcp-agent — Python client for LangGraph agents; load data360://agent-recipe for multi-agent nodes. Package README
LangChain minimal example — one-shot agent with data360://system-prompt. Example
LangGraph multi-agent example — gated tool node with relevance check. Example
@data360/mcp-ui — React components for MCP tool output (Vega charts, search results). Package README
@data360/mcp-ui-angular — Angular counterpart for host-embeddable chart UI. Package README
@data360/mcp-viz-core — Shared Vega-Lite prep and World Bank chart theme. Package README
@data360/tool-types — Zod schemas and parsers for MCP tool JSON contracts. Package README

Connect your agent

Required for external MCP clients (Cursor, Claude Desktop, LangGraph, and similar). Follow the steps below to wire your client to the public endpoint or a self-hosted instance.

1. Choose an endpoint

Use streamable HTTP at the /mcp path (default). For SSE, set MCP_TRANSPORT=sse on the server and point clients at /sse (example: http://localhost:8021/sse). Other org-specific hosts and paths are documented in the GitHub repository.

Environment	HTTP URL
Public (external)	`https://maimcpext.worldbank.org/ext/data360/mcp` Recommended for Cursor, Claude Desktop, and custom agents.
Local dev	`http://localhost:8000/mcp` Port matches your server (see README).

2. Configure your MCP client

Add the server to your client config. Streamable HTTP (type: "http") is the recommended transport for Cursor and compatible hosts.

{
  "mcpServers": {
    "data360-mcp": {
      "url": "https://maimcpext.worldbank.org/ext/data360/mcp",
      "type": "http"
    }
  }
}

{
  "mcpServers": {
    "data360-mcp": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://maimcpext.worldbank.org/ext/data360/mcp"
      ]
    }
  }
}

{
  "mcpServers": {
    "data360-mcp": {
      "url": "http://localhost:8000/mcp",
      "type": "http"
    }
  }
}

3. Load MCP resources in your agent

Load data360://system-prompt in your agent system context (required for reliable tool use).
Load data360://context for current date/year (recommended).
Optionally load data360://search-usage, data360://codelists, or data360://agent-recipe for LangGraph hosts.
Tables: search → get_disaggregation → get_data (or aggregation tools). Charts: search → get_disaggregation → get_viz_spec / get_multi_indicator_viz_spec.

4. Authentication (hosted / APIM)

World Bank–hosted endpoints may require a Bearer token or Azure AD client credentials. On the server, set MCP_INTERNAL=true and MCP_AUTH_SCOPE for APIM; clients may pass an Authorization header in mcp.json when required. See DEVELOPMENT.md for environment-specific setup.

5. Verify the connection

Use the MCP Inspector or your IDE's MCP panel to list tools. From the repo: scripts/test_mcp_client.py (data360-mcp on GitHub).

6. Docker and containers

From a container, use host.docker.internal instead of localhost to reach an MCP server on the host machine.

7. LangGraph / Python agents

Set environment variables before running your agent (see data360-mcp-agent package in the repo):

DATA360_MCP_URL=https://maimcpext.worldbank.org/ext/data360/mcp
DATA360_MCP_TRANSPORT=streamable_http
OPENAI_API_KEY=your-key

8. Self-host locally

Install and run the server on your machine: see the GitHub README.

Databases (examples)

All Data360 databases are supported; list indicators with data360_list_indicators.

WB_WDI — World Development Indicators
WB_SSGD — Social Sustainability and Global Database
WB_POVERTY — Poverty and inequality indicators
IPC_IPC — International Poverty Comparison

Health and operations

The FastAPI wrapper exposes HTTP probes for deployment and load balancers. See DEVELOPMENT.md for environment variables and configuration.

Route	Purpose	Status
`GET /mcp/health`	Liveness — process is running (no outbound I/O)	Always `200`
`GET /mcp/ready`	Readiness — Data360 API, database mapping, and viz storage are usable	`200` or `503`

Toggle readiness checks with MCP_READINESS_ENABLED (default true). Per-check timeout: MCP_HEALTH_CHECK_TIMEOUT (default 5 seconds).

World Bank development data through Data360 MCP