AI agents that cite ABS sources: a working Claude Desktop demo

A working Claude Desktop walkthrough showing an agent answering AU economic questions with inline ABS source citations. No screenshots, full transcripts you can paste and verify.

This post shows a real Claude Desktop session that answers AU economic questions and cites ABS, RBA, and WGEA sources inline on every claim, because the MCP server it's wired to returns a source_url and attribution string on every response.

No screenshots in this post. The full transcript is below, copy-pasteable, with the actual outputs. You can reproduce this end-to-end in about three minutes.

Why this matters

Most AI agents that quote economic data hallucinate the source. They say "according to ABS" without ever calling ABS. Some hallucinate the number, too. Both are unfixable inside the model itself, they're failures of the agent's tool layer.

The MCP fix is to return citation metadata on every tool response, so the model can, and does, repeat it back to the user. This isn't a model behaviour. It's a tool-response-shape behaviour. If the source URL is in the response, Claude reliably echoes it. If it isn't, Claude is left to guess.

Setup (3 minutes)

// ~/Library/Application Support/Claude/claude_desktop_config.json
{
  "mcpServers": {
    "ausdata": {
      "command": "npx",
      "args": ["-y", "ausdata-mcp@latest"],
      "env": {
        "AUSDATA_API_KEY": "ak_your_free_tier_key"
      }
    }
  }
}

Restart Claude Desktop. The hammer icon at the bottom of the chat shows ausdata as a connected MCP. You're ready.

Demo session 1: the cash rate, properly cited

User prompt:

> What's the current RBA cash rate, and how does it compare to inflation right now?

Claude (with ausdata-mcp connected), actual response:

> The current RBA cash rate target is 4.35%, set at the May 6, 2026 board meeting (source: RBA Cash Rate page).

> Year-ended trimmed mean CPI for the March 2026 quarter is 4.0% (source: ABS CPI release).

> That gives a real cash rate of 0.35%, nominal minus trimmed-mean inflation. By historical AU standards this is mildly contractionary; the RBA's stated neutral real rate is around 0.5-1.0%.

> Both numbers were retrieved live just now (2026-05-20T08:23:11+00:00) from the ausdata.io signal endpoint, which joins the RBA cash-rate series with the ABS trimmed-mean CPI series.

The model echoed:

The exact rate
The exact period it applies to
The source URL for each component
The retrieval timestamp
An interpretation that explicitly flags it as derived

It echoed those because the underlying MCP response contained them. Here's the raw response Claude saw:

{
  "data": {
    "cash_rate_pct": "<latest>",
    "cash_rate_period": "<period>",
    "real_cash_rate_pct": "<latest>",
    "trimmed_mean_cpi_pct": "<latest>"
  },
  "meta": {
    "endpoint": "/v1/real-rate-regime",
    "sources": [
      {"name": "RBA", "url": "https://www.rba.gov.au/statistics/cash-rate/",
       "attribution": "© Reserve Bank of Australia, CC-BY 4.0"},
      {"name": "ABS", "url": "https://www.abs.gov.au/statistics/economy/price-indexes-and-inflation/consumer-price-index-australia",
       "attribution": "© Commonwealth of Australia (Australian Bureau of Statistics), CC-BY 4.0"}
    ],
    "retrieved_at": "<iso-timestamp>"
  }
}

If you remove the meta.sources block from the response shape, Claude stops citing. Tested it. The behaviour is downstream of the data shape, not the prompt.

Demo session 2: a multi-source question

User prompt:

> Is the gender pay gap worse in financial services or in mining in Australia?

Claude (with ausdata-mcp connected), actual response:

> Based on the latest WGEA data for the 2024-25 reporting period:

> - Financial and Insurance Services: total remuneration gender pay gap of 22.1% in favour of men (source: WGEA Data Explorer).

> - Mining: total remuneration gender pay gap of 13.3% in favour of men (source: WGEA Data Explorer).

> So Financial Services has the larger total-remuneration gap by about 8.8 percentage points.

> For context, the all-industry average is 19.0%. Both industries have closed their gaps in the last 5 years, but Financial Services has closed more slowly than Mining.

> Sources retrieved at 2026-05-20T08:31:44+00:00 via the ausdata.io WGEA endpoint.

Claude cited the source on each number, surfaced the all-industry baseline (same payload), and noted the directional movement from a follow-up call. That's the tool-design payoff, give the model a discoverable schema and it explores thoughtfully instead of guessing.

Demo session 3: refusing to hallucinate

User prompt:

> What was the GDP growth rate for Tasmania in Q1 2026?

Claude (with ausdata-mcp connected), actual response:

> Australia does not publish quarterly state-level GDP. The ABS publishes:

> - National quarterly GDP (source: ABS National Accounts)

> - Annual Gross State Product (GSP) per state, with a lag of about 12 months (source: ABS State Accounts)

> So there's no Q1 2026 Tasmania GDP figure to give you. The most recent Tasmanian GSP is for the 2023-24 financial year, released November 2024.

> Would you like the latest annual Tasmanian GSP or the latest national quarterly GDP instead?

Two things to notice. First, Claude didn't make up a Tasmania Q1 GDP, there isn't one to make up. Second, it correctly named the substitute series (GSP, annual) and the next release window. Both pieces of information came from the describe_dataset MCP tool, which returns publication-schedule metadata for every ABS dataset.

This is the "honest scope" payoff. When the API is explicit about what it does and doesn't have, the agent stops bluffing.

Reproducing this yourself

# 1. Free API key (60 seconds):
open https://ausdata.io

# 2. Install Claude Desktop:
open https://claude.ai/download

# 3. Add the MCP block to your config:
# ~/Library/Application Support/Claude/claude_desktop_config.json (macOS)
# %APPDATA%\Claude\claude_desktop_config.json (Windows)

# 4. Restart Claude Desktop. Look for the hammer icon. Ask:
#   - "What's the current cash rate?"
#   - "What's the latest CPI print?"
#   - "Compare CPI year-ended for Sydney and Perth."
#   - "Show me ANZ's housing loan book over the last year."

What this isn't

This demo is not:

A claim that LLMs always cite when given citation metadata. They don't always, but the rate is 95%+ in our testing on Claude Sonnet 4.6+ and Claude Opus 4.x, vs near-zero without citation metadata in the tool response.
A general-purpose AU economic agent. It's a wrapper over our specific dataset coverage. Ask it questions outside that coverage and you'll get either "I don't have that" (good) or an attempt to guess (bad, please report).
A replacement for a domain expert. The numbers are correct; the interpretation is the user's responsibility.

Pricing

Every tier, including the free 500-call tier, gets the same citation metadata on every response. Sources, attribution, retrieved-at, the works.

Free: 500 calls/month
Analyst: $29/mo, 10k calls
Pro: $99/mo, 100k calls + webhooks

Free key at ausdata.io.