FAOSTAT MCP

FAOSTAT MCP Server

Query UN food and agriculture statistics with AI — powered by the Model Context Protocol

[Version] [PyPI] [MCP Registry] [Python 3.10+] [MCP Compatible] [License: MIT]

An MCP (Model Context Protocol) server that exposes the full FAOSTAT API as tools for AI assistants. Connect any MCP-compatible client — Claude, Cursor, Windsurf, Zed, or your own agent — to the world's most comprehensive database of food, agriculture, fisheries, forestry, and nutrition statistics, covering 245 countries and territories from the United Nations Food and Agriculture Organization (FAO).

Keywords: FAOSTAT, MCP server, Model Context Protocol, AI agriculture data, FAO statistics, food security AI, agricultural data Python, UN data, crop production statistics, Claude, Cursor, Windsurf

Why Use This?

Researchers, data journalists, policy analysts, and developers can ask natural-language questions and get answers directly from FAOSTAT — without writing a single API call. Your AI assistant handles domain discovery, filtering, and interpretation automatically.

Who is this for?

• Agricultural economists and food security researchers
• Journalists and policy analysts working with FAO data
• Developers building AI pipelines on top of FAOSTAT
• Anyone who wants to explore crop, trade, nutrition, or emissions data conversationally

What is FAOSTAT?

FAOSTAT is the statistical database of the United Nations Food and Agriculture Organization (FAO). It is the world's most comprehensive freely available source of data on food and agriculture, covering:

• Crop and livestock production — yields, harvested area, and quantities for hundreds of commodities
• Trade — import/export volumes and values between countries
• Food security — prevalence of undernourishment, dietary energy supply, and access indicators
• Emissions — greenhouse gas emissions from agriculture, land use, and food systems
• Forestry and fisheries — production and trade data
• Prices, inputs, and population — producer prices, fertilizer use, and demographic context

Data spans from 1961 to the present, across 245 countries and territories, in multiple languages.

What is MCP?

The Model Context Protocol is an open standard that lets AI assistants call external tools at runtime. This server registers all FAOSTAT API endpoints as discoverable tools — your AI assistant automatically selects and chains the right calls when you ask a question.

Features

• 23 MCP tools covering every FAOSTAT endpoint (data, metadata, rankings, bulk downloads, reports)
• 245 countries and territories across dozens of domains: crops, livestock, trade, food security, emissions, forestry, fisheries, and more
• Built-in rate limiting (2 req/s) — safe for the FAOSTAT production API out of the box
• Auto-retry with exponential backoff on transient network errors
• Rich tool descriptions so the AI knows exactly when and how to call each tool
• 3-tier hybrid caching — in-memory (20 min) → SQLite disk (24 h, cross-session) → Redis (optional, 30 min)
• Zero-config auth via faostat_setup — store credentials once, never touch a config file again
• Disambiguation via faostat_search_codes — agents ask before guessing ambiguous codes
• Works with Claude Desktop, Claude Code, Cursor, Windsurf, Zed, and any MCP-compatible client

Quick Start

Prerequisites

• Python 3.10+
• Any MCP-compatible client (Claude Desktop, Cursor, Windsurf, Zed, or a custom agent)

Option A — Install via MCP Registry (recommended)

Listed on the official MCP Registry — discoverable directly from Claude Desktop, Cursor, and any MCP-compatible client.

# Install with pip or uvx (no virtual env needed):pip install faostat-mcpuvx faostat-mcp

Updating: uvx faostat-mcp picks up new releases when your AI client restarts the server; if you still see an old version, run uvx faostat-mcp@latest once. pip and uv tool install don't upgrade on their own: run pip install -U faostat-mcp or uv tool upgrade faostat-mcp, then restart your client. The server also checks PyPI once per session and logs a notice to stderr when a newer version exists (opt out with FAOSTAT_NO_UPDATE_CHECK=1).

Option B — Install from source

git clone https://github.com/berba-q/faostat-mcp.gitcd faostat-mcppip install -e .

Configure credentials

Easiest — use the faostat_setup tool (no config files needed):

Once the server is running and connected to your AI client, ask your assistant:

"Call faostat_setup with my FAOSTAT username and password."

The tool validates your credentials against the API, then stores them securely in your system keychain (macOS/Windows) or ~/.config/faostat-mcp/credentials.json (Linux/Docker). All subsequent sessions authenticate automatically — no env vars or .env file required.

Alternative — environment variables (CI/CD, Docker, advanced):

cp .env.example .env# Edit .env:# FAOSTAT_USERNAME=your_email              ← recommended: tokens refresh automatically# FAOSTAT_PASSWORD=your_password# FAOSTAT_API_TOKEN=your_token_here        ← alternative: expires after 1 hour

Register for a free FAOSTAT API account at the FAOSTAT Developer Portal.

Optional — Redis caching (multi-user / high-volume deployments)

The server works without Redis (SQLite disk cache is used instead). For shared or high-volume setups, launch Redis via Docker:

docker run -p 6379:6379 -it redis/redis-stack:latest

Then set REDIS_HOST_IP_ADDRESS, REDIS_HOST_PORT_NUMBER, and REDIS_DATABASE in .env.

Running the Server

Development mode (interactive MCP Inspector UI)

mcp dev faostat_mcp/server.py

Opens a browser UI at http://localhost:5173 where you can browse and test all 23 tools interactively.

Production mode (stdio transport, for Claude Desktop)

python -m faostat_mcp.server# or, using the installed script:faostat-mcp

Caching

The server uses a 3-tier cache to minimise redundant API calls. FAOSTAT data updates at most daily, so most repeated queries are served instantly.

Cache lookup order: memory → disk → Redis → API call. A disk or Redis hit promotes the value to memory for the rest of the session.

To disable the disk cache (e.g. on a read-only filesystem), set FAOSTAT_DISK_CACHE=false.

MCP Client Integration

The server speaks standard MCP over stdio, so it works with any compatible client.

Recommended config (PyPI install)

{  "mcpServers": {    "faostat": {      "command": "uvx",      "args": ["faostat-mcp"]    }  }}

Dev / source config

{  "mcpServers": {    "faostat": {      "command": "python",      "args": ["-m", "faostat_mcp.server"],      "cwd": "/path/to/faostat-mcp",      "env": {        "FAOSTAT_USERNAME": "your_email",        "FAOSTAT_PASSWORD": "your_password"      }    }  }}

Claude Desktop

Add one of the blocks above to:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

Restart Claude Desktop — faostat will appear in the tools panel.

Cursor

Add the block to .cursor/mcp.json in your project root, or to your global Cursor MCP settings. See the Cursor MCP docs for details.

Windsurf / Zed / other clients

Any client that supports MCP stdio servers accepts the same config shape. Consult your client's documentation for the config file location.

Example Queries

Once connected, ask your AI assistant questions like:

Your AI assistant will automatically:

1. Call faostat_list_groups or faostat_groups_and_domains to find the right domain
2. Call faostat_search_codes to look up a code by name — if multiple codes match (e.g. "production" matches both Production and Gross Production Index), the assistant pauses and asks you to choose before proceeding
   • Regions and indicators are checked with faostat_resolve_name — if no matching FAOSTAT definition is found for your request (e.g. "Global South"), the assistant tells you so and offers the FAO-defined alternatives (e.g. "West Africa" → Western Africa) instead of inventing one
3. Call faostat_get_data or faostat_get_rankings with the confirmed codes
4. Interpret and summarize the results in plain language

Available MCP Tools

Discovery & Metadata

Data Retrieval

Authentication

Project Structure

faostat-mcp/├── pyproject.toml├── smithery.yaml             ← Smithery MCP registry manifest├── .env.example├── mcp_config_example.json   ← AI config snippet└── faostat_mcp/    ├── server.py             ← FastMCP server + all 23 tool definitions    └── client.py             ← HTTP client, rate limiting, 3-tier cache, credential storage

Important: Filter Codes vs Display Codes

The FAOSTAT API uses two different code systems: filter codes (used in query parameters) and display codes (shown in response data and bulk CSVs). Always use filter codes from faostat_get_codes when calling faostat_get_data.

Area, item, and year codes are the same for both. Only element codes differ:

QCL — Crops and Livestock Products

TM — Trade Matrix

FS — Food Security

Always call faostat_get_codes(dimension_id='element', domain_code=...) before querying data. Filter codes vary by domain and cannot be inferred from display codes.

# WRONG — uses display code 5510, returns empty datafaostat_get_data('QCL', area='2', item='515', element='5510', year='2024')
# CORRECT — uses filter code 2510, returns datafaostat_get_data('QCL', area='2', item='515', element='2510', year='2024')

Limitations & Notes

• This server targets the FAOSTAT production API (https://faostatservices.fao.org/api/v1).
• Rate limit: 2 requests/second, enforced automatically via token bucket.
• Responses are cached across 3 tiers (memory → SQLite disk → Redis) to reduce API calls — see .env.example for TTL and size configuration.
• The SQLite disk cache lives at ~/.cache/faostat-mcp/cache.db and defaults to 24 h TTL with a 1,000-entry LRU cap. Set FAOSTAT_DISK_CACHE=false to disable.
• For large domains (e.g., Trade Matrix), always apply area, item, and year filters to keep response sizes manageable.

Skills

Want guided analysis workflows on top of this server? Check out FAOSTAT Skills — 9 platform-agnostic AI skills for country profiles, commodity briefings, trade analysis, climate assessments, data visualization, and more. Works with Claude Code, OpenAI Codex, and any AI assistant that supports the SKILL.md format.

Related Projects & Resources

• FAOSTAT Skills — analysis skills built on top of this MCP server
• Model Context Protocol — the open standard powering this server
• FAOSTAT — UN FAO's official statistics portal
• FAOSTAT API Docs — developer reference
• Claude Desktop — one of the AI assistants this server works with
• Cursor — AI code editor with MCP support
• Windsurf — AI IDE with MCP support

Contributors

Thanks to everyone who has contributed to this project.

Contributions are welcome — see CONTRIBUTING.md for guidelines.

Citation

If you use this tool in academic work or research, please cite it:

Plain text:

Obli-Laryea, G., & Contributors. (2026). FAOSTAT MCP Server: AI-assisted access to FAOSTAT (v1.2.2) [Computer software]. https://github.com/berba-q/faostat-mcp

BibTeX:

@software{faostat_mcp,  author  = {Obli-Laryea, Griffiths and {Contributors}},  title   = {FAOSTAT MCP Server: AI-assisted access to UN food and agriculture statistics},  year    = {2026},  url     = {https://github.com/berba-q/faostat-mcp},  version = {1.2.2}}

See the Contributors section for a full list of authors.

When citing the underlying FAOSTAT data, use the FAO's recommended format with the specific domain:

FAO, {year}. FAOSTAT: {Domain Name}, http://www.fao.org/faostat/en/#data/{domain_code}

For example:

FAO, 2026. FAOSTAT: Crops and Livestock Products, http://www.fao.org/faostat/en/#data/QCL

FAO, 2026. FAOSTAT: Emissions Totals, http://www.fao.org/faostat/en/#data/GT

Changelog

See CHANGELOG.md for a full history of changes, generated automatically from conventional commits.

GitHub Topics

If you fork or star this repo, suggested topics: mcp, faostat, model-context-protocol, ai-tools, agriculture, food-security, fao, un-data, python, llm, unfao, undata

Smaller tool responses

Data, rankings, and global definitions default to response_format="compact" (column names once, followed by row arrays). Use response_format="objects" for the previous object-per-row shape, or "csv" for data and rankings. Select fields on data queries while preserving units needed to interpret values. Search results default to 25 matches; narrow the query if _truncated is true.

When the domain is known, use faostat_search_codes directly. The global name resolver returns definition_code, which is not necessarily a query filter code. A no_matching_definition result describes this lookup, not all FAO terminology. Code browsing, search, and validation share a cached table; repeated data queries reuse raw responses across formats, field selections, and row limits.