abs-mcp

Ask Claude about the Australian economy and get real, current numbers — not "I don't have access to that data." This MCP server gives Claude (and other MCP clients like Cursor) live access to the ABS Data API, with curated mappings for the 10 most-asked Australian economic indicators.

Behind the scenes it wraps SDMX 2.1, but you never see SDMX codes — just plain-English filters like region: "nsw" and measure: "unemployment_rate". Five tools, ten curated dataflows (Labour Force, CPI, Wage Price Index, Job Vacancies, Average Weekly Earnings, GDP / National Accounts, quarterly + annual Estimated Resident Population, Building Approvals, Lending Indicators), and 1,200+ other ABS dataflows accessible via raw codes.

Companion to rba-mcp (Reserve Bank of Australia — cash rate, FX, lending rates). Install both for the AU macro stack.

What you can ask

Once installed, your LLM can answer questions like:

Question	Real response (verified)
What's the unemployment rate in NSW?	4.27% (Mar 2026)
AU annual CPI inflation?	4.60% (Mar 2026)
AU annual wage growth?	3.40% (Q4 2025)
Average weekly earnings in Australia?	$1,562 (Sep–Oct 2025)
AU GDP quarterly growth?	0.80% (Q4 2025)
AU GDP per capita?	$24,900/qtr (Q4 2025)
Job vacancies in NSW?	101,200 (Q1 2026)
Dwelling approvals in NSW?	4,400/month (Mar 2026)
New NSW housing loan commitments?	$19.7B (Q4 2025)
Quarterly population of Australia?	27.7M (Q3 2025)

Every answer comes with the period, units, and a link back to the ABS source page. Comparisons and time-series queries work just as well — see Worked examples below.

Install

# After publish:
uvx abs-mcp

# Local dev install:
uv pip install -e .

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

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

For a local checkout (before PyPI publish):

{
  "mcpServers": {
    "abs": {
      "command": "uv",
      "args": ["run", "--directory", "/absolute/path/to/abs-mcp", "abs-mcp"]
    }
  }
}

Restart Claude Desktop. The abs server appears in the tools panel with five tools.

Cursor

Add to ~/.cursor/mcp.json (or workspace .cursor/mcp.json):

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

Tools

Tool	What it does
search_datasets(query, limit=10)	Fuzzy-search ABS dataflow names. Returns the top matches.
describe_dataset(dataset_id)	Plain-English description of a dataflow's dimensions and values.
get_data(dataset_id, filters, start_period, end_period, format)	Query a dataflow with filters. Returns clean records (default), grouped series, or CSV.
latest(dataset_id, filters)	Just the most recent observation(s) — wraps get_data with lastNObservations=1.
list_curated()	The ten dataflow IDs that have hand-curated plain-English support.

Curated dataflows

For these ten, filters accepts plain-English values (e.g. "region": "nsw" instead of "REGION": "1"):

• LF — Labour Force, monthly: employment, unemployment, participation by state/sex
• CPI — Consumer Price Index, quarterly inflation by capital city and category
• WPI — Wage Price Index, quarterly wage growth by industry/sector/state
• JV — Job Vacancies, quarterly labour demand by industry/sector/state
• AWE — Average Weekly Earnings, half-yearly by industry/sector/state
• ANA_AGG — National Accounts: GDP, GDP per capita, terms of trade, real income (Australia, quarterly)
• ABS_ANNUAL_ERP_ASGS2021 — Estimated Resident Population, annual by state and sub-state geography
• ERP_Q — Quarterly Estimated Resident Population, by state/sex/age
• BA_GCCSA — Building Approvals, monthly by state/capital region and building type
• LEND_HOUSING — Lending Indicators, quarterly new housing loan commitments by purpose, lender, and state

Any other ABS dataflow still works — pass raw SDMX dimension IDs and codes.

Worked examples

"What's the current unemployment rate in NSW?"

Claude calls:

latest(dataset_id="LF", filters={"region": "nsw", "measure": "unemployment_rate"})

Returns:

{
  "dataset_id": "LF",
  "dataset_name": "Labour Force",
  "query": {"region": "nsw", "measure": "unemployment_rate"},
  "period": {"start": "2026-03", "end": "2026-03"},
  "unit": "Percent",
  "records": [
    {
      "period": "2026-03",
      "value": 4.27,
      "dimensions": {"measure": "Unemployment rate", "region": "New South Wales", "sex": "Persons"},
      "unit": "Percent"
    }
  ],
  "source": "Australian Bureau of Statistics",
  "retrieved_at": "2026-05-11T03:14:22Z",
  "abs_url": "https://www.abs.gov.au/statistics/labour/employment-and-unemployment/labour-force-australia"
}

"Show me NSW housing approvals over the last two years"

get_data(dataset_id="BA_GCCSA", filters={"region": "nsw", "measure": "dwelling_units"}, start_period="2024")

"Compare quarterly CPI in Sydney vs Melbourne"

get_data(dataset_id="CPI", filters={"region": ["sydney", "melbourne"], "measure": "change_year"}, start_period="2023")

Period formats

ABS uses different period formats per dataflow. Pass start_period / end_period in the matching format:

Dataflows	Frequency	Format	Example
LF, BA_GCCSA	Monthly	YYYY-MM	"2026-03"
CPI, WPI, JV, ANA_AGG, LEND_HOUSING, ERP_Q	Quarterly	YYYY-Q* or YYYY-MM	"2025-Q4"
AWE	Half-yearly	YYYY-S*	"2025-S2"
ABS_ANNUAL_ERP_ASGS2021	Annual	YYYY	"2025"

Development

git clone https://github.com/Bigred97/abs-mcp.git
cd abs-mcp
uv sync --extra dev
uv pip install -e .

# Unit tests (no network)
uv run pytest

# Live integration tests (hits real ABS API)
uv run pytest -m live

The SQLite cache lives at ~/.abs-mcp/cache.db. Catalogue refreshes every 24h, codelists every 7 days, data responses every hour, latest 15 minutes. Delete the file to force a refresh.

How it works

When you ask Claude an ABS question, it picks the right tool, fills in the curated filters, and calls the live ABS API. You see the reasoning + tool call inline:

Claude does the picking; this server does the SDMX translation, unit attribution, and clean response shaping. You don't have to know what M13.3.1599.20.1.M means — and neither does Claude.

How it differs from existing ABS MCP servers

The one existing community option (seansoreilly/abs) exposes a single query_dataset tool that passes raw SDMX through. This package offers semantic tools and curated mappings for the highest-value dataflows so an LLM can answer real questions without you needing to know what M13.3.1599.20.1.M means.

Changelog

See CHANGELOG.md for release history.

License

MIT — Harry Vass, 2026.