oecd-mcp-server

mcp
Guvenlik Denetimi
Uyari
Health Uyari
  • License — License: Apache-2.0
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Low visibility — Only 5 GitHub stars
Code Gecti
  • Code scan — Scanned 12 files during light audit, no dangerous patterns found
Permissions Gecti
  • Permissions — No dangerous permissions requested

Bu listing icin henuz AI raporu yok.

SUMMARY

Search, explore, and query 1,500+ OECD statistical datasets (national accounts, employment, trade, education, health) from the OECD SDMX API via MCP. STDIO or Streamable HTTP.

README.md

@cyanheads/oecd-mcp-server

Search, explore, and query 1,500+ OECD statistical datasets (national accounts, employment, trade, education, health) from the OECD SDMX API via MCP. STDIO or Streamable HTTP.

8 Tools • 1 Resource • 1 Skill

Version License Docker MCP SDK npm TypeScript Bun

Install in Claude Desktop Install in Cursor Install in VS Code

Framework

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


Overview

OECD statistics across 1,500+ datasets on national accounts, employment, trade, education, and health. Search datasets, inspect their dimensions, query observations, and run SQL over large multi-country time series. No API key required.

Tools

Five discovery and data tools plus three DataCanvas tools for large query results:

Tool Description
oecd_list_agencies List OECD SDMX agencies with their directorate and the number of dataflows each publishes
oecd_search_datasets Search 1,500+ OECD dataflows by keyword or theme
oecd_get_dataset_info Fetch a dataflow's dimensions, key order, and codelist references
oecd_get_dimension_values Fetch valid codes and labels for one dimension (countries, measures, frequencies)
oecd_query_dataset Fetch observations filtered by dimension key and time range; spills large results to DataCanvas
oecd_dataframe_describe List DataCanvas tables and columns staged by a prior oecd_query_dataset spill
oecd_dataframe_query Run a read-only SQL SELECT against DataCanvas tables
oecd_dataframe_drop Drop a staged DataCanvas table or view once the analysis is done — opt-in via OECD_DATAFRAME_DROP_ENABLED=true

Resources

Resource Description
oecd://dataflow/{agency_id}/{flow_id} Dimension metadata for a single OECD dataflow — same content as oecd_get_dataset_info

All resource data is also reachable via tools. Use oecd_get_dataset_info for the same content.

Skill

Skill Description
oecd-guide Building SDMX keys, confirming which series a dataflow carries, and reading results without mixing breakdowns, units, or country aggregates. Served over MCP at skill://oecd-guide/SKILL.md

Capability reference

oecd_list_agencies tool

  • No inputs; returns every agency ID (e.g. OECD.SDD.NAD) with its directorate name and dataflow count, sorted by count
  • Publishers outside OECD that ship dataflows through the same catalog (ESTAT, IAEG-SDGs) carry no directorate

oecd_search_datasets tool

  • Each query token must start a word in a dataflow's name or description: matching is a case-insensitive prefix, and a word starts after any character that is not a letter or digit, or at a lowercase-to-uppercase step (markers matches RioMarkers). A token that opens on such a character, like & or -oecd, matches anywhere it occurs, so & finds R&D. tiva finds (TiVA) but not activation, and employment does not reach a flow that only mentions unemployment
  • Results rank in three tiers: name matches first, then tokens split across name and description, then description-only, in catalog order within each tier. Optional agency_id scope
  • limit defaults to 20 and takes any value up to 2,000; a larger one is served at 2,000 with a notice. offset pages the ranked list, and every response carries total_matches
  • Each result carries its flow_ref — the identifier every other tool takes — and matched_in (name, description, or both)

oecd_get_dataset_info tool

  • Takes a flow_ref; returns every dimension in key order with its concept name and codelist reference, plus a key_example for oecd_query_dataset
  • Fails invalid_flow_ref for a string that is not a flow reference and dataflow_not_found for one OECD does not publish

oecd_get_dimension_values tool

  • Takes a flow_ref and dimension_id; optional query substring-matches both code and label
  • A codelist is shared across dataflows and can list codes a flow holds no data for. available_only lists only the codes with data, from OECD's availability record; key (the oecd_query_dataset form) lists the codes with data given the key's other segments, and implies available_only. Either adds time_range, observation_count, and available_only: true to the result. An availability read that fails falls back to the whole codelist with available_only: false and a notice naming the failure. A key holding a character outside the SDMX key set, or with too few segments, fails invalid_key
  • Availability costs one structure request, sent alongside the codelist read, and no data download. A delegated dataflow's availability is read on the root its data lives on, the one its catalog entry points to. A cold call sends 3 requests without the flag (dataflow reference, concept scheme, codelist) and 4 with it, or 5 for a delegated dataflow whose datastructure /public/rest also holds, which adds the datastructure route; a repeat call is served from the structure cache with no request
  • limit (default 50) and offset page the matches, with the full match count when more remain; a limit above 10,000 is served at 10,000 with a notice

oecd_query_dataset tool

  • flow_ref plus a dot-delimited key (+ for multiple values, empty segments as wildcards) and optional start_period / end_period
  • One row per observation, sorted by dimension code in key order and then by TIME_PERIOD. A row carries value (pre-multiplied by value_scale) and each component that varies across the result; a coded component appears twice, as its label and as <ID>_code
  • Components holding one value across the whole result come back once, in constant_columns ({ code, label } per component; {} when nothing is constant), and leave the rows. Attributes are read at dataset, dimension-group, and observation level, and UNIT_MULT scales the value from whichever level carries it
  • A query that matches nothing fails no_results, naming the pinned code that has no data, a code matching only by letter case, or a period outside the key's range. Producing that costs one structure request once the dataflow's structure is cached (three on a cold call: dataflow reference, concept scheme, availability) and no data download; if the read fails or takes longer than 10 seconds, the plain message stands
  • Large results spill to DataCanvas with truncated: true, canvas_id, and table_name; without a canvas every row stays in structuredContent and only the rendered table is capped (content_table_capped)
  • A spill includes a notice naming the canvas and table: inspect columns with oecd_dataframe_describe, then use oecd_dataframe_query for SQL over the full result. The staged table holds every column, constants and _code columns included, so SQL can filter on codes. Fitting results allocate no new canvas; an explicit canvas_id is checked before the download, so a stale one fails as canvas_not_found without spending an OECD request.

oecd_dataframe_describe tool

  • Takes the canvas_id oecd_query_dataset returned; lists each staged table and view with its row count and column names and types

oecd_dataframe_query tool

  • canvas_id plus one read-only SQL SELECT; writes, DDL, and system-catalog access are rejected
  • limit (default 100) returns up to 1,000 rows per call; a larger limit is served at 1,000 with a notice
  • truncated: true marks a result with more rows than it returned, and its notice names how to reach the rest; row_count counts the rows returned
  • A query still running after 30 seconds is stopped with query_timeout

oecd_dataframe_drop tool

  • canvas_id plus a table_name as oecd_dataframe_describe lists it; returns the dropped object's kind and the remaining_tables
  • Opt-in: listed but not callable unless OECD_DATAFRAME_DROP_ENABLED=true

oecd://dataflow/{agency_id}/{flow_id} resource

  • Same content as oecd_get_dataset_info, as application/json
  • {flow_id} is {dsd_id}@{df_id} with @ encoded as %40 (e.g. oecd://dataflow/OECD.SDD.NAD/DSD_NAAG%40DF_NAAG_I), or the bare {df_id}

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.

OECD-specific:

  • Keyless access — no API key required; OECD SDMX 2.1 REST API is fully public
  • Covers 1,500+ dataflows across 20+ OECD statistical departments (national accounts, employment, inflation, trade, education, health, environment, taxation, inequality)
  • Delegated dataflows and codelist revisions resolved end to end — a dataflow OECD catalogues on one service root but defines on another follows the catalog's own link for structure, codes, and observations, and codes come from the revision the datastructure names rather than the endpoint's current latest
  • Structure responses (the dataflow catalog, datastructures, codelists, concept schemes, availability) are cached in process for the lifetime OECD grants each, bounded at 64 MiB, so repeating a lookup inside that window sends no OECD request. The bound and lifetime are not configurable; observations are never cached
  • AllDimensions observation mode — one-pass SDMX-JSON decoding into flat row objects, no nested series key reconstruction
  • oecd_query_dataset materializes large observation sets (multi-country time-series) on a DuckDB DataCanvas for in-conversation SQL analytics; the oecd_dataframe_* tools work on it; DataCanvas is on by default and CANVAS_PROVIDER_TYPE=none disables it

Agent-friendly output:

  • Workflow-aware tool surface — flow_ref from search flows directly into info, values, and query tools without reconstruction
  • Spill signaling — truncated: true + canvas_id tells the agent to switch to SQL instead of parsing a truncated inline list
  • Full SDMX decoding server-side — agents see { REF_AREA: "United States", REF_AREA_code: "USA", MEASURE: "Gross domestic product", MEASURE_code: "B1GQ", value: 26054614000000, value_scale: 1000000000 }, with components that never vary hoisted into constant_columns, not raw index arrays

Getting started

Connect to the hosted instance, no install:

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

Or run it locally:

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

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

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

No API key is needed — OECD SDMX is a free, public API. DataCanvas SQL analytics over large query results is on by default; set CANVAS_PROVIDER_TYPE=none to get capped inline previews instead. For Docker or a source build, see Running the server.

Configuration

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

Variable Description Default
OECD_BASE_URL OECD SDMX REST API base URL. Must be an https origin that answers directly — no redirect is followed, so a plaintext http:// origin fails instead of being upgraded to https. https://sdmx.oecd.org/public/rest
OECD_TIMEOUT_MS Per-request timeout in milliseconds. Each tool or resource call also spends one upstream budget of max(50 s, this value) across all its requests and retries, so a value past 50 s needs a client that waits that long for a result. 30000
OECD_CLAMP_LIMITS Serve a limit above a tool's per-call cap (oecd_search_datasets 2,000, oecd_get_dimension_values 10,000, oecd_dataframe_query 1,000) at the cap, with a notice. false rejects it instead and advertises the cap as the schema maximum. true
CANVAS_PROVIDER_TYPE Canvas engine. With duckdb, a large oecd_query_dataset result spills to a queryable table instead of just capping the rendered preview; with none, every row still comes back in structuredContent and only the rendered table is capped. duckdb
OECD_DATAFRAME_DROP_ENABLED Set to true to register oecd_dataframe_drop. Off, the tool is listed as disabled and clients cannot call it. false
MCP_TRANSPORT_TYPE Transport: stdio or http. stdio
MCP_HTTP_PORT Port for HTTP server. 3010
MCP_SESSION_MODE HTTP session posture: stateful, stateless, or auto. The server declares stateless in source — it keeps no per-session state, and a DataCanvas handle is keyed by canvas_id — so set this only to override that. stateless
MCP_AUTH_MODE Auth mode: none, jwt, or oauth. none
MCP_LOG_LEVEL Log level (RFC 5424). info
LOGS_DIR Directory for log files (Node.js only). <project-root>/logs
LOG_TOOL_FAILURE_PAYLOADS Log each failed tool call's arguments and result, redacted by key name and capped at LOG_TOOL_FAILURE_PAYLOAD_MAX_BYTES (default 16384). A secret inside a free-form value is not redacted. false
OTEL_ENABLED Enable OpenTelemetry instrumentation. false

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

Running the server

Local development

Requires Bun v1.4.4 or higher (or Node.js v24+).

  • Clone and install:

    git clone https://github.com/cyanheads/oecd-mcp-server.git
    cd oecd-mcp-server && bun install
    cp .env.example .env   # optional — every variable has a default
    
  • 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 test suite
    bun run lint:mcp   # Validate MCP definitions against spec
    

Docker

docker run --rm -p 3010:3010 ghcr.io/cyanheads/oecd-mcp-server:latest

The image defaults to HTTP transport, stateless session mode, and logs to /var/log/oecd-mcp-server. To run it over stdio from an MCP client, use "command": "docker" with "args": ["run", "-i", "--rm", "-e", "MCP_TRANSPORT_TYPE=stdio", "ghcr.io/cyanheads/oecd-mcp-server:latest"]. To build locally, docker build -t oecd-mcp-server .; OpenTelemetry peer dependencies are installed by default — build with --build-arg OTEL_ENABLED=false to omit them.

Project structure

Directory Purpose
src/index.ts createApp() entry point — registers tools/resources and initializes services.
src/config/ Server-specific environment variable parsing and validation with Zod.
src/mcp-server/tools/definitions/ Tool definitions (*.tool.ts) — eight tools for OECD data discovery, retrieval, and DataCanvas analysis.
src/mcp-server/resources/definitions/ Resource definitions (*.resource.ts) — the oecd://dataflow resource.
src/services/oecd-http/ Shared OECD fetch boundary — timeout and retry-classification corrections used by both services below, the response-body reader both share, the per-call upstream budget, the origin check every delegated service root passes before it is addressed, the refusal of any redirect off the configured host, and the classification that gives an upstream refusal the same declared reason on every tool and resource.
src/services/oecd-structure/ OECD SDMX structure service — dataflows, data structures, codelists, availability — and the in-process cache in front of its reads.
src/services/oecd-data/ OECD SDMX data service — observations and SDMX-JSON decoding.
src/services/canvas-accessor/ DataCanvas accessor — registers and exposes the framework canvas instance to tools.
src/utils/ Shared helpers — table formatting, limit clamping, word-start text matching.
skills/ Published agent skill (oecd-guide), served over MCP and shipped in the npm package and Docker image.
tests/ Unit and integration tests mirroring src/.

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 request-scoped logging, ctx.state for tenant-scoped storage
  • Register new tools and resources via the barrels in src/mcp-server/*/index.ts
  • Wrap external API calls: validate raw SDMX-JSON → normalize to domain type → return output schema; never fabricate missing fields

Contributing

Issues are welcome — see CONTRIBUTING.md for what makes one actionable, and CODE_OF_CONDUCT.md for how we work together. Run checks and tests before submitting:

bun run devcheck
bun run test

License

Apache-2.0 — see LICENSE for details.

Yorumlar (0)

Sonuc bulunamadi