ISTAT MCP Server - Workflow Guide

Language Detection — First Step (mandatory)

Before doing anything else, detect the language of the user's first message.

User writes in	Set	Apply to
Italian	`lang=it`	all responses, labels, summaries, warnings, `get_concepts` calls, "Fonti dati" section
English	`lang=en`	all responses, labels, summaries, warnings, `get_concepts` calls, "Data Sources" section

Rules:

Detect once at the start; keep lang fixed for the entire conversation even if the user later switches language.
Every response — including tool-call summaries, warnings, and the closing data-sources section — must use the detected language.
When calling get_concepts, always pass "lang": "<detected>" (e.g. {"concept_id": "FREQ", "lang": "it"}).
If the language cannot be determined (e.g. the first message is a bare dataflow ID), default to lang=it.

Quick Start

If the query targets a specific territory (region, province, municipality), start by resolving the territory:

Resolve territory with get_territorial_codes → get REF_AREA codes
Discover the dataflow with discover_dataflows
Get metadata with get_constraints (one call returns dimensions + valid codes + descriptions)
Fetch data with get_data using the codes from steps 0 and 2

Skip step 0 only when the query is about Italy as a whole (REF_AREA: IT).

Time Period Rule

Default behaviour — last available year, month or quarter. Unless the user explicitly requests a time series or a specific period, always use get_data without start_period/end_period and with lastNObservations=1 . Using lastNObservations=1 returns only the last available observation. This keeps responses compact and avoids flooding the context with multi-decade series.

Time series — explicit request. Use start_period / end_period only when the user explicitly asks for:

time series for a specific years range ("dal 2015 al 2026", "from 2010 to 2026")
Time series starting from a specific year ("dal 2015 in poi", "from 2015 onwards")
time series for a specific month/quarter ("mostrami la serie storica mensile", "show me the quarterly series"). In this case show data only for the last 2 available years to avoid flooding the context with too many observations.

In that case omit both start_period and end_period entirely — the server fetches all years available for that dataflow based on the TIME_PERIOD constraint.

Quick decision:

User asks "quali sono i dati sulla disoccupazione?" → no periods but lastNObservations=1

User asks "mostrami la serie storica dal 2010 al 2026" → set start_period + end_period

User asks "mostrami la serie storica dal 2015" → set start_period only, leave end_period empty

User asks "dammi la serie storica completa" → omit both start_period and end_period

output

Always show the url used for 'get_data' and the dataflow name, so the user can replicate the query independently. See "Fonti dati / Data Sources" section below for the required format. See Generate Download URL for the full workflow.

Available Tools

#	Tool	Purpose
1	`discover_dataflows`	Find datasets by keywords (with blacklist filtering)
2	`get_constraints`	Retrieve constraints + structure + descriptions in one call
3	`get_structure`	Retrieve dimensions and codelists definitions
4	`get_codelist_description`	Retrieve IT/EN descriptions for codelist values
5	`get_concepts`	Retrieve the IT or EN description of an SDMX concept by ID — always pass `lang` detected at conversation start
6	`get_data`	Retrieve statistical observations
7	`get_cache_diagnostics`	Debug tool to inspect cache status
8	`get_territorial_codes`	Lookup REF_AREA codes by level, name, region, province, or capoluogo

Fast Path: Skip get_constraints when get_constraints fails

When you already know the codes to use (e.g. common values like FREQ=A, SEX=9, REF_AREA from get_territorial_codes) and only need the dimension order, you can use get_structure passing id_datastructure associated with the id_dataflow from discover_dataflows instead get_constraints:

When to use this:

You know the code values but not the dimension order
You want to skip the full get_constraints call (which fetches codelists for every dimension)
You're doing exploratory queries with wildcards (omit unknown dimensions from dimension_filters)

When to still use get_constraints:

You don't know which codes are valid for a dimension
You need human-readable descriptions of codes

Detailed Workflow

Step 0: Resolve Territory

Tool: get_territorial_codes

Use this tool before querying data whenever the user mentions a specific place, area, or territorial grouping. Never guess REF_AREA codes — always resolve them through this tool.

When to use it:

User mentions a region, province, or municipality by name (e.g. "dati sulla Campania", "province del Veneto")
User asks about a group of territories (e.g. "regioni del Sud", "capoluoghi di provincia")
User needs to compare territories (e.g. "Nord vs Sud vs Centro")
User asks about comuni with specific characteristics (e.g. "capoluoghi della Lombardia")
The dataflow has a REF_AREA dimension and the query is not about Italy as a whole

Examples:

[
  { "level": "regione" },
  { "name": "Milano" },
  { "level": "provincia", "region": "Sicilia" },
  { "level": "comune", "region": "Lombardia", "capoluogo": true }
]

The tool contains 9,142 entries with the full Italian territorial hierarchy (italia → ripartizione → regione → provincia → comune) and parent-child relationships. It also flags capoluoghi di provincia and di regione.

Use the returned codes directly in the REF_AREA dimension filter of get_data.

IF the user query is about Italy as a whole, skip this step and use REF_AREA: IT in get_data.

Step 1: Identify Dataflows

Use discover_dataflows with comma-separated keywords (Italian or English).

{ "keywords": "employment,labour,work" }

Output: list of dataflows with ID, names (IT/EN), and descriptions.

Note: dataflows in the blacklist (env var DATAFLOW_BLACKLIST) are automatically excluded.

Choosing the right dataflow when multiple candidates exist

When discover_dataflows returns more than one matching dataflow, always prefer the one with the most recent data. To compare them:

Call get_constraints on each candidate (results are cached, so only the first call hits the API).
Read the EndPeriod value of the TIME_PERIOD dimension.
Select the dataflow with the highest (most recent) EndPeriod.

If two dataflows have the same EndPeriod, prefer the one whose name or description better matches the user's intent.

Example: if the results include DF_DCIS_DISOCCUPATI_1 (EndPeriod 2022) and DF_DCIS_DISOCCUPATI_2 (EndPeriod 2025), always use DF_DCIS_DISOCCUPATI_2.

Step 2: Retrieve Constraints and Descriptions

Use get_constraints with the dataflow ID from step 1. This is the recommended approach - one call instead of many.

{ "dataflow_id": "101_1015_DF_DCSP_COLTIVAZIONI_1" }

This internally calls get_structure + get_codelist_description for each dimension. Everything is cached for 1 month.

Output includes:

Dimension names and order
Valid codes for each dimension (only values available for that dataflow)
Italian and English descriptions for each code
Codelist IDs

Typical output:

{
  "id_dataflow": "101_1015_DF_DCSP_COLTIVAZIONI_1",
  "constraints": [
    {
      "dimension": "FREQ",
      "codelist": "CL_FREQ",
      "values": [
        {"code": "A", "description_en": "Annual", "description_it": "Annuale"}
      ]
    },
    {
      "dimension": "TYPE_OF_CROP",
      "codelist": "CL_AGRI_MADRE",
      "values": [
        {"code": "APPLE", "description_en": "Apples", "description_it": "Mele"},
        {"code": "WHEAT", "description_en": "Wheat", "description_it": "Grano"}
      ]
    },
    {
      "dimension": "TIME_PERIOD",
      "StartPeriod": "2006-01-01T00:00:00",
      "EndPeriod": "2026-12-31T23:59:59"
    }
  ]
}

Alternative manual approach: call get_structure first, then get_codelist_description for each codelist you need.

To explore the values of a specific codelist:

{
  "codelist_id": "CL_ATECO_2007"
}

Optional: Understand SDMX Concepts

Tool: get_concepts

Use this tool to understand the semantics of the dataflow's concepts (dimensions and attributes) when the meaning of a dimension is unclear from get_constraints output alone. Always pass the lang detected at conversation start:

{ "concept_id": "AGRIT_AUTHORIZATION", "lang": "it" }

The tool returns a single description string in the requested language.

Step 3: Fetch Data

Tool: get_data

This tool makes the final call to the ISTAT endpoint to retrieve observations.

Rules for building filters

Time periods — default: last available year. Do NOT set start_period/end_period unless the user explicitly requests a historical series or a specific range. When omitted, get_data automatically selects the last available year from TIME_PERIOD constraints. Set periods only when the user asks for historical data or a specific range.
Dimension order: The order of filters must match the one returned by get_constraints.
Dimensions without filter: Use . to indicate “all values”.
Multiple filters on a dimension: Concatenate codes with + (OR operator).

Use get_data with dimension filters built from step 2 output.

{
  "id_dataflow": "149_577_DF_DCSC_OROS_1_1",
  "dimension_filters": {
    "FREQ": ["Q"],
    "REF_AREA": ["IT"],
    "DATA_TYPE": ["FT_EMPL_1"],
    "ADJUSTMENT": ["N"],
    "ECON_ACTIVITY_NACE_2007": ["0011", "0013", "0015"]
  },
  "start_period": "2020-01-01",
  "end_period": "2023-12-31"
}

Query examples

Query 1 — Monthly historical series:

{
  "id_dataflow": "22_315_DF_DCIS_POPORESBIL1_2",
  "dimension_filters": {
    "FREQ": ["M"],
    "REF_AREA": ["IT"],
    "DATA_TYPE": ["DEROTHREAS"],
    "SEX": ["9"]
  },
  "start_period": "2019-01-01",
  "end_period": "2025-11-30",
  "detail": "full"
}

Query 2 — Quarterly data with sector filter:

{
  "id_dataflow": "149_577_DF_DCSC_OROS_1_1",
  "dimension_filters": {
    "FREQ": ["Q"],
    "REF_AREA": ["."],
    "DATA_TYPE": ["."],
    "ADJUSTMENT": ["."],
    "ECON_ACTIVITY_NACE_2007": ["0011", "0013", "0015"]
  },
  "start_period": "2020-09-01",
  "end_period": "2023-12-31",
  "detail": "full"
}

Key rules for get_data:

Dimension order must follow the order from get_constraints
Multiple codes for the same dimension: use an array ["0011", "0013"]
No filter on a dimension: omit it from dimension_filters
Default behavior: if no time range is specified, only the latest available year is returned
Rate limit: the ISTAT API allows max 3 calls per minute (handled automatically)

Complete Use Case: Employment by Sector

Scenario

Analyze employment in Italian manufacturing sectors from 2020 to 2023.

Step 1 — Find the dataflow

{
  "keywords": "occupazione,ore,lavorate"
}

→ We identify 149_577_DF_DCSC_OROS_1_1.

Step 2 — Get constraints

{
  "dataflow_id": "149_577_DF_DCSC_OROS_1_1"
}

→ We obtain dimensions, valid codes and time range.

Step 3 — Retrieve data

{
  "id_dataflow": "149_577_DF_DCSC_OROS_1_1",
  "dimension_filters": {
    "FREQ": ["Q"],
    "REF_AREA": ["IT"],
    "DATA_TYPE": ["FT_EMPL_1"],
    "ADJUSTMENT": ["N"],
    "ECON_ACTIVITY_NACE_2007": ["0011", "0013", "0015"]
  },
  "start_period": "2020-01-01",
  "end_period": "2023-12-31",
  "detail": "full"
}

Generate Download URL (skip get_data)

When the user asks for a download URL, a link to the data, or says "dammi l'URL per scaricare…", do NOT call get_data. Instead, build the URL yourself from the constraints metadata. This avoids dumping large datasets into the conversation context.

When to use this mode

User says "URL", "link", "scarica", "download" or similar
User explicitly does not want data displayed, just a way to get it
The goal is to hand off the download to the user (browser, curl, DuckDB, script)

Workflow

Follow steps 0–2 of the standard workflow (resolve territory → discover dataflow → get constraints), then:

Step 3 (URL-only): Build the URL from constraints

Use the dimension order and codes from get_constraints to construct the SDMX URL.

URL pattern:

https://esploradati.istat.it/SDMXWS/rest/data/{dataflow_id}/{dim1.dim2.dim3...}/ALL/?startPeriod={start}&endPeriod={end}&format=csv

Rules for building the dimension path:

Dimensions must appear in the exact order from get_constraints
Filtered dimension: codes joined with + (e.g., ITC1+ITC2+ITC3)
Unfiltered dimension (all values): leave empty (just the . separator)
Every dimension must be represented, even if empty

Example: unemployment rate by region, last 5 years

After get_constraints on dataflow 151_914, the dimension order is: FREQ.REF_AREA.DATA_TYPE.SEX.AGE.EDU_LEV_HIGHEST.CITIZENSHIP.DURATION_UNEMPLOYMENT

Build the path:

Dimension	User intent	Filter
FREQ	annual	`A`
REF_AREA	all 21 regions	`ITC1+ITC2+ITC3+ITC4+ITD1+ITD2+ITD3+ITD4+ITD5+ITE1+ITE2+ITE3+ITE4+ITF1+ITF2+ITF3+ITF4+ITF5+ITF6+ITG1+ITG2`
DATA_TYPE	unemployment rate	`UNEM_R`
SEX	total	`9`
AGE	15-64	`Y15-64`
EDU_LEV_HIGHEST	total	`99`
CITIZENSHIP	total	`TOTAL`
DURATION_UNEMPLOYMENT	total	`TOTAL`

Resulting URL:

https://esploradati.istat.it/SDMXWS/rest/data/151_914/A.ITC1+ITC2+ITC3+ITC4+ITD1+ITD2+ITD3+ITD4+ITD5+ITE1+ITE2+ITE3+ITE4+ITF1+ITF2+ITF3+ITF4+ITF5+ITF6+ITG1+ITG2.UNEM_R.9.Y15-64.99.TOTAL.TOTAL/ALL/?startPeriod=2021&endPeriod=2026&format=csv

Output format

Respond with:

CSV URL — ready to open in browser or use with curl/DuckDB
curl command — curl "URL" for terminal download
Query summary — dataflow name, dimensions used, period, any choices made
Warnings — if a requested breakdown doesn't exist, explain what was used instead

Example response:

URL CSV (apri nel browser o scarica con curl):
https://esploradati.istat.it/SDMXWS/rest/data/151_914/A.ITC1+ITC2+...TOTAL/ALL/?startPeriod=2021&endPeriod=2026&format=csv

curl:
curl "https://esploradati.istat.it/SDMXWS/rest/data/151_914/A.ITC1+ITC2+...TOTAL/ALL/?startPeriod=2021&endPeriod=2026&format=csv"

Dettagli:
- Dataflow: 151_914 — Tasso di disoccupazione
- Frequenza: annuale
- Territorio: tutte le 21 regioni italiane
- Fascia d'età: 15-64
- Sesso: totale
- Periodo: 2021-2026

Tips

For DuckDB users, suggest: SELECT * FROM read_csv_auto('URL');
If the dimension path becomes very long, the URL is still valid — SDMX handles long query strings
Always validate codes against get_constraints output before building the URL
Use get_territorial_codes to resolve territory names to REF_AREA codes — never guess

Visualization

Age pyramid:

The age pyramid display is only used when displaying the population by gender and age group for a given year. The pyramid should be displayed so that the highest age values are at the top and the age group starting at zero years at the bottom.

Best Practices

Always use get_constraints before get_data to know the correct dimension order and available codes.
Default: last available year — never set start_period/end_period unless the user explicitly asks for a historical series or a specific time range. Omitting them lets the server auto-select the last available year, keeping the response compact.
Historical series — only when explicitly requested: set start_period and end_period from the TIME_PERIOD constraint (or leave both empty to fetch all available years).
Start with few filters and add more progressively to avoid empty datasets.
Dimensions without filter: always represent with ., never omit.
Multiple filters: concatenate with + (e.g. "ECON_ACTIVITY_NACE_2007": ["0011", "0013"]).
Inspect codelist values to pick exact, valid codes
Cache is your friend: metadata cached 1 month · dataflows 7 days · observed data 24h (1 day)
get_data is lean by design: it reads only raw constraints from cache (dimension order + TIME_PERIOD range) without loading codelist descriptions. Codelist descriptions are only fetched when get_constraints is called explicitly. The processed TSV is cached directly, so cache hits return the ready-to-use table with no re-parsing.
Never guess REF_AREA codes: always use get_territorial_codes to resolve place names to codes. Territory is often the starting point of any ISTAT query.
Always show data sources: after every get_data call, close with a "Fonti dati" / "Data Sources" section that includes: dataflow ID + name, CSV URL and curl command (from get_data output → "How to reproduce this query"), filters per dimension, and period used. Never omit the curl command — it lets the user replicate the query independently.

Troubleshooting

Problem	Solution
Dataset too large	Add dimensional filters or reduce the time range
No data returned	Verify that codes exist in the codelist and are compatible with each other
Wrong dimension order	Check `get_constraints` output for the correct order
Malformed query string (404)	Empty dimensions must be `.`; when there is a filter, `.` still follows
Don't know the REF_AREA code	Use `get_territorial_codes` with name or level to find the right code
Don't know the dimensions of a dataflow	Two fast calls (~1.5s total): 1) get structure ID: `curl -s "https://esploradati.istat.it/SDMXWS/rest/dataflow/IT1/{id}"` → grep for `Ref id`; 2) get dimensions in order: `curl -s "https://esploradati.istat.it/SDMXWS/rest/datastructure/IT1/{struct_id}"` → grep for `Dimension`
Error 500 on a dataflow ID	The ID may be a parent container (e.g. `39_493`). Use `discover_dataflows` to find the sub-dataflows (e.g. `39_493_DF_DCIS_CMORTE1_EV_1`)

Output finale / Data Sources

Dopo ogni analisi che usa get_data, chiudi SEMPRE con una sezione il cui titolo dipende dalla lingua rilevata:

lang=it → "Fonti dati"
lang=en → "Data Sources"

Cosa mostrare

L'output di get_data include già in coda una sezione "How to reproduce this query" con:

CSV URL (apribile nel browser)
SDMX URL (XML grezzo)
comando curl per scaricare il CSV
breakdown dei filtri per dimensione

Non ricostruire queste informazioni da zero. Riportale direttamente dalla sezione "How to reproduce this query" presente nell'output di get_data, aggiungendo solo le informazioni mancanti elencate sotto.

Formato obbligatorio

Adatta l'intestazione alla lingua rilevata (lang=it → "Fonti dati", lang=en → "Data Sources").

## Fonti dati          ← oppure "Data Sources" se lang=en

**Dataflow**: `{id_dataflow}` — {nome_it} / {nome_en}
{descrizione breve, se disponibile da discover_dataflows}

**Query** *(from get_data output)*:

```
{csv_url}
```

```bash
curl "{csv_url}"
```

**Filtri applicati** *(dimension order from get_constraints)*:
- `{DIM1}`: `{codice/i}`
- `{DIM2}`: `{codice/i}` — *(all values)* se non filtrata
- ...
- Periodo: `{start_period}` → `{end_period}` oppure "ultimo anno disponibile ({anno})"

**Campi nel risultato**: {lista colonne TSV, separata da virgole}
{Eventuali avvertenze, es. "Nessun dato per alcune combinazioni di filtri"}

Regole

Dataflow sempre nominato: mostra sempre id_dataflow + nome (IT se lang=it, EN se lang=en). Se disponibile dalla risposta di discover_dataflows, includi anche la descrizione.
CSV URL e curl sempre presenti: copiati dall'output di get_data → sezione "How to reproduce this query". Non ricostruire l'URL a mano.
Filtri sempre esplicitati: lista ogni dimensione con i codici usati. Se una dimensione non è filtrata, scrivi (all values).
Periodo sempre indicato: se start_period/end_period sono stati passati, mostrali; altrimenti scrivi "ultimo anno disponibile" con l'anno effettivo leggibile dal TSV.
Più dataflow: se la risposta combina più chiamate get_data, ripeti il blocco per ciascun dataflow.

API Reference

Base URL: https://esploradati.istat.it/SDMXWS/rest
Format: SDMX 2.1 XML → TSV output
Rate Limit: 3 calls/minute (automatically managed by the MCP server)
Cache: metadata 1 month · dataflow 7 days · observed data 24h (1 day)
Query path format: /data/{dataflow_id}/{dim1.dim2.dim3...}/ALL/?params
- Empty dimensions: .
- Multiple values: + (e.g. IT+FR)
- All dimensions must be present in the order from get_structure

istat-mcp