Tools

Four tools and one resource. You rarely call them directly — the assistant chains them; this page is the map.

The server is deliberately small: a discovery pair, a resolver, a fetcher, and a guide. They're meant to be used in order.

​ search_datasets → inspect_dataset → resolve_series → query_data find the see its real turn filters fetch the dataflow codes & coverage into one key observations ​

A full round-trip, in one line each: search "GDP per capita" to find the dataflow, inspect it to see which countries and measures actually have data, resolve {country: Portugal, measure: GDP per capita} into the exact positional key, then query that key for the numbers, unit, and source link.

search_datasets

Find an OECD dataflow by topic keywords. Returns dataflow ids and names.

The entry point when you don't already know which dataset holds what you want. One topic can span several dataflows (house prices, rents, and price-to-income may live in different ones), so it returns candidates rather than a single guess.

inspect_dataset

Given a dataflow, returns its dimensions in positional order, and for each, the codes that actually have data (with human labels) — not the full theoretical codelist. Also returns the real time coverage, observation count, and the dataflow's description.

This is where you choose codes deliberately: it surfaces the ~20 countries that have data rather than the 469 the codelist allows, so you filter on what's real.

resolve_series

The core tool. Turns dimension filters given by name (country, measure, frequency…) into one exact, validated series key — you never hand-build a positional key. It:

• rejects codes that have no data, before any data call (unavailable_value),
• auto-fills any dimension with a single possible value,
• narrows to the real series, and if the filters still match several, returns exactly which dimensions disambiguate them (ambiguous_filters),
• returns a terminal no_series_exist when a combination genuinely has no data — a definitive stop, not a retry hint.

The failure reasons are described on Honest failures.

query_data

Fetches the observations for a resolved key. Alongside the data, every result carries what a quoted figure needs:

• source.api — the OECD API URL the numbers came from,
• source.dataExplorer — a human, clickable OECD Data Explorer link to the exact series, to verify and download,
• measurement — the unit, price base, transformation, and adjustment, so the figure can't be restated wrong (current-PPP dollars, not euros).

An empty result here is terminal too (series_has_no_data), so a valid-looking but dataless key can't become an invitation to guess.

Resource: oecd://guide

A read-only usage guide that teaches a model the workflow above, how to disambiguate, and the honesty rules. Load it into the system prompt for clients that don't read MCP resources automatically — see Connect it. It's the difference between a model that follows the intended flow and one that improvises.

For exact shapes, read the source

Each tool's inputs and outputs are described in its own schema at runtime. This page is a map, not a regenerated parameter dump — for precise types, the tool source is the source of truth.