faostat-mcp-server-216ddf— plugin

<div align="center"> <h1>@cyanheads/faostat-mcp-server</h1> <p><b>Global food & agriculture statistics from the UN FAOSTAT bulk-download corpus, served from a local SQLite mirror with a DataCanvas SQL surface, via MCP. STDIO or Streamable HTTP.</b> <div>7 Tools • 1 Skill Resource • 0 Prompts</div> </p> </div>

<div align="center">

</div>

<div align="center">

Public Hosted Server: https://faostat.caseyjhand.com/mcp

</div>

Overview

Global food and agriculture statistics from the UN's FAOSTAT: production, trade, food balances, food security, land use, fertilizer, and emissions for 245+ countries and territories since 1961. Resolve commodity and country codes, query the data, and run SQL over large results.

Tools

Capability reference

faostat_list_domains _tool

• Full FAOSTAT catalog (~69 domains) read live from the bulk manifest, annotated with local mirror status
• code for an exact domain lookup; topic substring filter over code/name/topic; indexed_only to list only domains queryable from the local mirror
• offset + limit (max 200, default 20) page the catalog — response reports totalMatches, truncated, and nextOffset
• Each entry reports indexed / index_ready flags, local row count, and last completed sync

faostat_resolve_codes _tool

• FTS5 full-text query, substring name_contains, or exact code lookup within a dimension: area, item, or element
• Item/element matches are scoped to codes present in the given domain's cube; area codes are shared across domains
• Every area match is flagged country or aggregate (codes ≥ 5000, plus curated sub-threshold roll-ups such as China=351)
• limit (max 200, default 50) + offset page the match set
• Typed errors: unknown_domain, index_not_ready, query_timeout

faostat_query_observations _tool

• Filters by area_codes / item_codes / element_codes and an inclusive year_start / year_end range
• Aggregate regions excluded by default (include_aggregates: false); explicit area_codes bypass the exclusion
• limit caps the inline page (default 200, max 1000); a match that exceeds it spills in full to a DataCanvas table (50,000-row staging cap) — faostat_dataframe_describe for its columns, then SQL via faostat_dataframe_query
• Every row carries its data-quality flag (A/E/I/B/M/T/X, others per domain) — never dropped
• Typed errors: domain_not_indexed, index_not_ready, canvas_disabled, canvas_not_found, invalid_year_range, query_timeout

faostat_commodity_profile _tool

• Selects one QCL item with item_code or item_query (code takes precedence), then ranks top producers/exporters/importers and returns an annual production trend
• Query selection prefers a unique exact label after trimming, collapsing whitespace, and case folding; otherwise requires a unique relevance match. For ambiguous terms such as maize or milk, call faostat_resolve_codes with domain: "QCL", dimension: "item", and the query, then pass the selected item_code (e.g. 56 for maize grain)
• Rankings are per-country sums grouped by unit, each country at its own latest reporting year; countries only (aggregates excluded)
• Returns a partial, production-only profile with a notice — rather than failing — when the trade domain (TCL) isn't indexed or still syncing
• top_n caps each ranked list (max 50); when the merged observation set is too large to inline it spills to a DataCanvas table — faostat_dataframe_describe for its columns, then faostat_dataframe_query for further SQL
• Every staging outcome is stated in the notice: the staged table, a set that fit inline, a staging failure, or a disabled canvas
• Typed errors: missing_item, ambiguous_item, no_match, domain_not_indexed, index_not_ready, invalid_year_range, canvas_not_found, query_timeout

faostat_dataframe_query _tool

• Single-statement read-only SELECT over staged faostat_xxxxxxxx tables — joins, aggregates, window functions, and CTEs all work
• Writes, DDL, DROP, COPY, PRAGMA, ATTACH, external-file functions, and system catalogs (information_schema, sqlite_master, duckdb_*) are rejected
• Engine-setting introspection, query_table, dynamic query() SQL, and file table references are rejected; query() with a single literal read-only SELECT remains supported
• row_limit caps the response (default 1000, max 10000); truncated means more rows exist, with no exact total computed on this path
• Typed errors: canvas_disabled, canvas_not_found, missing_table, system_catalog_access, invalid_sql

faostat_dataframe_describe _tool

• Lists staged tables with source tool, query params, row count, column schema, and recorded creation/expiry timestamps
• name describes one table outright; otherwise offset + limit (max 100, default 20) page the listing newest-first
• Typed errors: canvas_disabled, canvas_not_found, missing_table

faostat_dataframe_drop _tool

• Opt-in: registered disabled — listed on the landing page with its enable hint, absent from tools/list — unless FAOSTAT_DATAFRAME_DROP_ENABLED=true
• Drops one staged faostat_xxxxxxxx table and its provenance from the session canvas, or from the canvas a canvas_id names
• Idempotent: a name not staged on that canvas returns dropped: false with a notice and changes nothing
• Typed errors: canvas_disabled, canvas_not_found

Features

The faostat-guide covers comparable reporting years, country versus regional totals, units, coverage, and complete SQL analysis across staged results. It ships with plugin, npm, Docker, and MCPB installs and is served through the MCP Skills extension as a skill resource.

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.

FAOSTAT-specific:

• Persistent local SQLite mirror of the FAOSTAT bulk corpus via the framework MirrorService, with FTS5 over the dimension labels driving code resolution
• Streaming bulk-ZIP ingester — skips domains whose upstream update date hasn't advanced, and stream-parses the normalized CSV into SQLite without materializing the full file in memory
• Config-driven domain selection (FAOSTAT_DOMAINS) — the indexed set can grow without code changes, and the full catalog stays browsable regardless
• DataCanvas SQL surface (DuckDB) for GROUP BY, ranking, and time-series analysis over spilled result sets

Agent-friendly output:

• Country-vs-aggregate classification on every area, with aggregates excluded from sums by default — guards against double-counting World/continent rows with their member countries
• Data-quality provenance — every observation carries its FAOSTAT flag (A/E/I/B/M/T/X, others per domain), never dropped from output
• Graceful partial results — faostat_commodity_profile returns a production-only profile with a notice, rather than failing, when the trade domain isn't indexed
• Typed error contracts — index_not_ready, domain_not_indexed, canvas_disabled, and others each carry a concrete recovery hint

Getting started

Connect to the hosted instance, no install:

claude mcp add --transport http faostat-mcp-server https://faostat.caseyjhand.com/mcp

Or run it locally after building the mirror:

claude mcp add faostat-mcp-server -- npx -y @cyanheads/faostat-mcp-server

For other clients, add this to the MCP config file:

{
  "mcpServers": {
    "faostat-mcp-server": {
      "command": "npx",
      "args": ["-y", "@cyanheads/faostat-mcp-server"]
    }
  }
}

The bulk-download service is public and keyless. For Docker or a source build, see Running the server.

Configuration

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

Running the server

Local development

Requires Bun v1.4.0 or higher (or Node.js v24+), plus disk for the mirror. The default domain set (QCL,TCL,FBS,FS,RL,GLE,RFN,QV, ∼37M rows) needs a few GB; drop TCL (∼17M rows) from FAOSTAT_DOMAINS on a constrained host.

git clone https://github.com/cyanheads/faostat-mcp-server.git
cd faostat-mcp-server && bun install
cp .env.example .env
# Override the default domain set, mirror path, or refresh cron in .env.

• Build and run:

  # One-time build
  bun run rebuild

  # Run the built server
  bun run start:stdio
  # or
  bun run start:http

• Run checks and tests:

  bun run devcheck   # Lint, format, typecheck, security
  bun run test       # Vitest suite under Bun, then under Node
  bun run lint:mcp   # Validate MCP definitions against spec

Building the mirror

The corpus is not bundled. Before the data tools can answer queries, sync the selected domains into the local mirror:

bun run mirror:init      # one-time bootstrap — downloads and indexes the FAOSTAT_DOMAINS set
bun run mirror:refresh   # re-sync domains whose upstream update date has advanced; build missing planner statistics
bun run mirror:verify    # report sync status, local row counts, and sample reads

mirror:init is idempotent and resumable per domain — re-running after an interrupt re-streams only the unfinished domain ZIP. FAOSTAT_DOMAINS selects which domains are indexed; everything else in the catalog shows in faostat_list_domains with indexed: false until added and re-synced. On HTTP transport, set FAOSTAT_REFRESH_CRON to refresh in-process on a schedule; on stdio, run mirror:refresh out-of-band. The read tools return index_not_ready until the first sync completes.

mirror:refresh also builds query-planner statistics (ANALYZE) for any selected domain that has none, even when its data is unchanged upstream — a mirror synced by an older release, or one whose last build failed. Otherwise the server builds them itself, in the background on a worker thread, the first time a data read touches that domain, and keeps answering meanwhile.

Docker

docker run --rm -p 3010:3010 -v faostat-mirror:/usr/src/app/.faostat-mirror ghcr.io/cyanheads/faostat-mcp-server:latest

Build locally with docker build -t faostat-mcp-server .. For stdio from an MCP client, use "command": "docker" with "args": ["run", "-i", "--rm", "-e", "MCP_TRANSPORT_TYPE=stdio", "-v", "faostat-mirror:/usr/src/app/.faostat-mirror", "ghcr.io/cyanheads/faostat-mcp-server:latest"]. The image defaults to HTTP transport, stateless session mode, and logs to /var/log/faostat-mcp-server. A dedicated dependency stage installs the production tree for the target platform (DuckDB's prebuilt native binding included), so the slim runtime image carries no build toolchain and a multi-arch build never runs Bun under emulation. OpenTelemetry peer dependencies are installed by default — build with --build-arg OTEL_ENABLED=false to omit them. Mount a volume at the mirror path to persist the corpus across container recreations, and bootstrap it inside the container:

docker exec <container> bun run mirror:init      # one-time bootstrap
docker exec <container> bun run mirror:verify    # sync status + sample reads
docker exec <container> bun run mirror:refresh   # re-sync updated domains; build missing planner statistics

Project structure

Development guide

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

• Handlers throw, framework catches — no try/catch in tool logic
• Use ctx.log for request-scoped logging, ctx.state for tenant-scoped storage
• Register new tools in the createApp() array in src/index.ts
• Wrap external data: validate raw → normalize to domain type → return output schema; never fabricate missing fields, and carry the data-quality flag through

Data attribution

Data is sourced from FAOSTAT, the statistics division of the Food and Agriculture Organization of the United Nations (FAO). FAOSTAT data is published under CC BY-4.0; cite FAO as the source in downstream use. This project is not affiliated with or endorsed by the FAO.

Contributing

Issues are welcome. Run checks and tests before submitting:

bun run devcheck
bun run test

License

Apache-2.0 — see LICENSE for details.