unhcr-refugees-mcp-server

---

Overview

Refugee and displacement statistics from the UNHCR Refugee Data Finder: refugee, asylum-seeker, IDP, and stateless populations since 1951, asylum applications and decisions since 2000, and returns and resettlement since 1959. Pull annual figures by country of origin and asylum, and run SQL over large results. No API key required.

Tools

Tool	Description
unhcr_list_reference	Decode countries (ISO3, ISO2, and UNHCR codes, names, regions), regional bureaus, dataset coverage years, population types, and asylum codes
unhcr_get_population	Year-end displacement stocks by origin and/or asylum country from 1951, with UNRWA and IDMC series alongside and an optional current-year nowcast
unhcr_get_demographics	Year-end stocks by population type, sex, and age band from 2001, with the share UNHCR could disaggregate
unhcr_get_asylum_applications	Asylum applications lodged per year from 2000, split by application stage by default
unhcr_get_asylum_decisions	Asylum decisions per year from 2000 by outcome, with the Refugee Recognition Rate and Total Protection Rate
unhcr_get_solutions	Refugee returns, resettlement, naturalisation, and IDP returns per year from 1959
unhcr_dataframe_describe	Describe a staged dataframe by name, or list them all where listing is on, with provenance, expiry, completeness, and column schema
unhcr_dataframe_query	Run one read-only SQL SELECT across staged dataframes, optionally saving the result as a new one
unhcr_dataframe_drop	Drop a staged dataframe before its TTL — opt-in, absent from tools/list by default

unhcr_dataframe_drop is registered only when UNHCR_DATAFRAME_DROP_ENABLED=true; the other eight tools are always advertised.

Dataframe workflow

1. A unhcr_get_* call whose full result exceeds limit, or that sets stage: true, stages every row as a df_XXXXX_XXXXX table and returns its handle in dataset.
2. unhcr_dataframe_describe lists the staged tables with their columns, source call, and expiry. Listing is off over HTTP with MCP_AUTH_MODE=none, so there describe each table by the name in dataset.
3. unhcr_dataframe_query runs one read-only DuckDB SELECT across them; register_as saves the result as a new table for the next query.
4. Tables expire after UNHCR_DATASET_TTL_SECONDS (24 hours by default). With UNHCR_DATAFRAME_DROP_ENABLED=true, unhcr_dataframe_drop removes one sooner. Once a tenant's staged rows pass 1,000,000, the oldest tables are evicted first, and the call that pushed the total over names them in evicted.

Dataframes run on DuckDB's native binding, which the npm and Docker installs carry. The Claude Desktop .mcpb bundle ships without it, so there the data tools answer inline only. Set CANVAS_PROVIDER_TYPE=none to turn dataframes off anywhere.

Capability reference

The five unhcr_get_* tools share one contract:

• Scope: origin and asylum take ISO3 codes, case-insensitive, as a string or a list, up to 50 each. ISO2 codes are rewritten to ISO3; UNHCR's own codes and country names are rejected. Each listed code returns its own rows, and an omitted dimension is summed into one row unless expand (origin, asylum, or both) lists every country. year_from / year_to default to the dataset's span and are clamped to it.
• Rows: the full result is sorted by sort_by, then cut to limit (default 100). Requests from 1 to 1,000,000 are accepted; values above 500 clamp to 500, with requested and effective limits disclosed even for an empty or small result. Where dataframes are on, the full set of a larger result is staged, and stage: true stages a result that fits too.
• Output: total_rows, complete (false when UNHCR_MAX_ROWS stopped the fetch), measure (stock or flow), applied_scope, latest_year, dataset when staged, data_notes, and attribution.
• Errors: unknown_country_code, invalid_year_window, year_out_of_coverage, conflicting_scope, and the retryable upstream_busy, which carries retryAfter.

---

unhcr_list_reference _tool

• topic: countries, regions, coverage, population_types, or asylum_codes. With countries, name_contains keeps countries whose names contain every word given, or whose ISO3, ISO2, or UNHCR code equals one; there is no fuzzy matching
• coverage gives each dataset's first_year, latest_year, and measure, plus the month of the current nowcast
• Countries, regions, and coverage come from the live API and are cached for 24 hours; population types and asylum codes are static, and asylum codes UNHCR's methodology does not define carry documented: false

---

unhcr_get_population _tool

• Year-end stocks refugees, asylum_seekers, oip, idps, stateless, ooc, and hst, plus returned_refugees and returned_idps, which are flows during the year; sort_by takes any of them or year
• unrwa_refugees (Palestine refugees registered with UNRWA) and idmc_conflict_idps (IDMC's conflict-IDP estimate) sit beside a row when that series has one and are never added into UNHCR's counts. Matching UNHCR footnotes come back too, up to 20, with footnotes_total
• include_nowcast: true appends UNHCR's current-year monthly estimate of refugees and asylum-seekers by asylum country; it is skipped when origin lists codes

---

unhcr_get_demographics _tool

• population_types filters to REF, ASY, OIP, IDP, STA, OOC, HST, RET, or RDP; sort_by takes year or total
• Each row is one population type: total, fourteen sex × age bands (female_0_4 … female_60_plus, female_unknown_age, female_total, and the male_* twins), disaggregated, and sex_disaggregated_share (0–1). Bands are null where UNHCR has no breakdown
• Totals come from a separate collection and need not match unhcr_get_population; matching footnotes come back as there

---

unhcr_get_asylum_applications _tool

• split_by names which of authority, stage, and decision_level stay separate rows (default ["stage"]; [] gives one total per year, scope, and unit). stages filters before summing, e.g. ["N"] for new applications; sort_by takes year or applied
• Each row lists the authorities, stages, and decision_levels codes summed into it, its unit (persons or cases), and applied. Cases are never added to persons

---

unhcr_get_asylum_decisions _tool

• split_by names which of authority and decision_level stay separate rows (default [], all summed, as UNHCR does for its rates). decision_levels filters before summing, e.g. ["FI"] for first instance
• Each row carries recognized, complementary_protection, rejected, otherwise_closed, total_decisions, substantive_decisions, and a unit, plus refugee_recognition_rate and total_protection_rate as percentages of substantive decisions. A rate is null when that denominator is 0 or null
• sort_by takes year or a count (total_decisions, substantive_decisions, recognized, rejected); rates are not sortable

---

unhcr_get_solutions _tool

• Flows returned_refugees, resettlement, naturalisation, and returned_idps; sort_by takes any of them or year. Matching footnotes come back as with population
• The asylum country means something different per column: the country refugees returned from, the country they were resettled to, the country that naturalised them. IDP returns sit on the origin country itself

---

unhcr_dataframe_describe _tool

• Optional name (df_XXXXX_XXXXX); omit it to list every staged table, newest first. Over HTTP with MCP_AUTH_MODE=none listing is off and name is required
• Each entry gives source_tool, query_params, created_at, expires_at, row_count, complete, incomplete_sources, providers, and column_schema. Saved descendants retain original incomplete-source handles after their ancestors expire or are evicted. Staged tables add origin_unhcr_region, origin_unsd_region, asylum_unhcr_region, and asylum_unsd_region columns the inline rows omit

---

unhcr_dataframe_query _tool

• sql is one DuckDB SELECT of at most 20,000 characters; joins, aggregates, window functions, and CTEs work. Writes, DDL, PRAGMA, ATTACH, file references, external-data functions, engine introspection (current_setting, current_schema, current_schemas, in_search_path), dynamic table lookup (query_table), and system catalogs are rejected with a typed reason. query() accepts only a single SQL string literal, checked by the same gate
• row_limit (1–10,000, default 1,000) caps the rows the query materializes, and row_count_capped flags a capped result. preview returns fewer rows inline; register_as saves the result under a new df_XXXXX_XXXXX name with a fresh TTL. It counts toward the 1,000,000-row staging budget, and a result that alone exceeds it fails register_as_too_large
• SUM and COUNT values come back as strings (BIGINT). attribution credits the third-party series of every table the SQL reads
• preview accepts 0–1,000,000 and clamps to row_limit, disclosing requested and effective values. Zero still returns no preview rows. complete and incomplete_sources describe upstream coverage independently of display truncation: aggregates, joins and saved descendants remain partial if an original source was partial
• The serialized tool result, including both response surfaces and notices, stays within 65,536 UTF-8 bytes. Whole rows or oversized metadata are withheld identically; values are never clipped. result_chunks identifies a tenant-scoped snapshot of the materialized columns, rows and result metadata. It has the usual TTL and shares the staging budget; result_storage_unavailable means that snapshot could not be retained. A saved register_as table remains available

For a result_chunks.name such as df_ABCDE_12345, read one chunk per call:

SELECT chunk_index,chunk FROM df_ABCDE_12345 WHERE chunk_index=0

Increment the index through result_chunks.chunk_count - 1. Base64-decode each chunk separately (at most 6,144 source bytes), concatenate the decoded bytes in index order, then decode UTF-8 and parse JSON. A byte boundary may split a Unicode character, so decode UTF-8 only after joining the bytes. The snapshot contains rows materialized up to row_limit, including rows hidden by preview; row_count_capped still means additional matching rows were not materialized. For the full registered result, query registered_as. Small scalar results use Markdown tables; values needing exact whitespace, control characters, nested structures or Markdown syntax use fenced JSON.

---

unhcr_dataframe_drop _tool

• name (df_XXXXX_XXXXX); returns dropped: false when nothing matched, so repeating a call is safe
• Registered only when UNHCR_DATAFRAME_DROP_ENABLED=true; otherwise the per-table TTL reclaims staged tables

Features

Built on @cyanheads/mcp-ts-core: stdio and Streamable HTTP transports, pluggable auth (none / jwt / oauth), swappable storage (in-memory, filesystem, Supabase, Cloudflare KV/R2/D1), structured logging with optional OpenTelemetry tracing.

UNHCR-specific:

• Keyless client for the API behind UNHCR's Refugee Data Finder (api.unhcr.org/population/v1). The API ignores unknown parameters and lone year bounds rather than rejecting them, so the server sends only allowlisted parameters, always pairs the year bounds, and pins ISO3 country matching (cf_type=ISO)
• Shared request queue: paced request starts (UNHCR_REQUESTS_PER_SECOND), at most two in flight, a cooldown after a 429, identical requests collapsed into one, a response cache, and a 45-second budget per call
• UNHCR's reading of the numbers: stocks and flows labelled as such, UNHCR's "-" returned as null rather than zero, cases never added to persons, recognition rates over UNHCR's substantive-decision denominator
• UNHCR rounds counts below 5 (below 10 for asylum decisions) to the nearest multiple of 5, so small values are approximate
• Every result carrying UNHCR figures cites the "UNHCR Refugee Population Statistics Database" (CC BY 4.0) with a link to UNHCR's Terms of Use for Datasets. UNRWA and IDMC figures are those providers' own series, carried through UNHCR's API; attribution.providers credits them whenever they appear, since their own conditions may apply

Agent-friendly output:

• Provenance on every data result: applied_scope echoes the codes, year window, and ISO2 rewrites actually sent, alongside latest_year, measure, complete, and attribution
• Caveats as data: data_notes say how to read each result (stock or flow, rounding, what null means), and matching UNHCR footnotes arrive with rows_matched
• Typed failures with recovery hints: unknown_country_code suggests the ISO3 for a UNHCR code, year_out_of_coverage names the dataset's span, upstream_busy carries retryAfter
• Nothing trimmed silently: year clamps, empty-result hints, truncation, and where the full set was staged all arrive as a notice in the response

Getting started

Public Hosted Instance

A public instance is available at https://unhcr-refugees.caseyjhand.com/mcp — no installation required. Point any MCP client at it via Streamable HTTP:

{
  "mcpServers": {
    "unhcr-refugees-mcp-server": {
      "type": "streamable-http",
      "url": "https://unhcr-refugees.caseyjhand.com/mcp"
    }
  }
}

Every caller of the hosted instance shares one UNHCR request queue and one dataframe workspace, so unhcr_dataframe_describe takes a table's exact name there; listing is off.

Self-Hosted / Local

Add the following to your MCP client configuration file.

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

Or with npx (no Bun required):

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

Or with Docker:

{
  "mcpServers": {
    "unhcr-refugees-mcp-server": {
      "type": "stdio",
      "command": "docker",
      "args": ["run", "-i", "--rm", "-e", "MCP_TRANSPORT_TYPE=stdio", "ghcr.io/cyanheads/unhcr-refugees-mcp-server:latest"]
    }
  }
}

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 v1.4.0 or higher (or Node.js v24+).
• No API key or account: UNHCR's API is keyless, and every variable below is optional.

Installation

1. Clone the repository:

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

2. Navigate into the directory:

cd unhcr-refugees-mcp-server

3. Install dependencies:

bun install

4. Configure environment:

cp .env.example .env
# optional: every variable has a working default

Configuration

Variable	Description	Default
UNHCR_REQUESTS_PER_SECOND	Upstream request starts per second for the whole process (1–10). Every caller of an HTTP instance shares it.	4
UNHCR_MAX_ROWS	Most upstream rows one tool call fetches (10,000–500,000). A larger result stops at the cap and reports complete: false.	150000
UNHCR_CACHE_MAX_MB	Response-cache budget in MB; 0 disables it. Reference data (countries, regions, coverage, footnotes) has its own cache, always on.	64
UNHCR_DATASET_TTL_SECONDS	Lifetime of a staged dataframe, in seconds (minimum 60).	86400
UNHCR_DATAFRAME_DROP_ENABLED	Set true to register unhcr_dataframe_drop.	false
CANVAS_PROVIDER_TYPE	Dataframe engine: duckdb, or none to turn dataframes off.	duckdb
MCP_TRANSPORT_TYPE	Transport: stdio or http.	stdio
MCP_HTTP_PORT	HTTP server port.	3010
MCP_SESSION_MODE	HTTP session mode: stateless, stateful, or auto.	stateless
MCP_AUTH_MODE	Authentication: none, jwt, or oauth.	none
MCP_LOG_LEVEL	Log level (debug, info, warning, error, etc.).	info
LOGS_DIR	Directory for log files (Node.js only).	<app-root>/logs
STORAGE_PROVIDER_TYPE	Storage backend: in-memory, filesystem, supabase, cloudflare-kv/r2/d1.	in-memory
OTEL_ENABLED	Enable OpenTelemetry.	false

See .env.example for the full list of optional overrides.

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
  bun run test:live  # Opt-in: re-checks the upstream behaviors the client relies on against api.unhcr.org
  

Project structure

Directory	Purpose
src/index.ts	createApp() entry point: server instructions, tool registration, service setup.
src/config	Server-specific environment variable parsing and validation with Zod.
src/mcp-server/tools/definitions	Tool definitions (*.tool.ts) and the registration list that gates unhcr_dataframe_drop.
src/mcp-server/tools/shared	Scope inputs, output schemas, result shaping, and markdown rendering shared by the data tools.
src/services/unhcr	UNHCR API client — request builder, paced fetch, response cache, reference data, normalization, asylum aggregation, footnote matching.
src/services/canvas-bridge	Adapter to the framework DataCanvas — dataframe naming, provenance, expiry, and the read-only SQL gate.
tests/	Unit tests mirroring the src/ structure, plus the opt-in live suite in tests/live/.

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 in buildToolDefinitions() in src/mcp-server/tools/definitions/index.ts
• Wrap external API calls: validate raw → normalize to domain type → return output schema; never fabricate missing fields

Contributing

Issues 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.

The data belongs to UNHCR and is published under CC BY 4.0 and UNHCR's Terms of Use for Datasets. This project is independent of UNHCR and not endorsed by it.