If you are building an AI agent on Claude (or Cursor, or Windsurf, or any MCP-compatible host) that needs current Australian macro, financial, or energy data, here is how to wire it up in 10 minutes.

The problem you are solving

Claude training data has a cutoff. If you ask "what is the current RBA cash rate" without grounding, Claude will either refuse to answer or guess from training data that is months stale. For agents handling real financial questions, this is unacceptable.

The Model Context Protocol (MCP) is Anthropic open standard for letting models call live tools. Install the ausdata MCP server and Claude can fetch live data from 9 Australian government sources as easily as it can read a file.

Install in 30 seconds

The ausdata MCP server is a single npm package:

npx ausdata-mcp

For Claude Desktop, add this to your config (Settings → Developer → Edit Config):

{
  "mcpServers": {
    "ausdata": {
      "command": "npx",
      "args": ["ausdata-mcp"],
      "env": { "AUSDATA_API_KEY": "ak_YOUR_KEY" }
    }
  }
}

Restart Claude Desktop. The ausdata tools appear in the tool list.

Cursor

In Cursor settings, MCP Servers → Add new:

{
  "name": "ausdata",
  "command": "npx",
  "args": ["ausdata-mcp"],
  "env": { "AUSDATA_API_KEY": "ak_YOUR_KEY" }
}

Windsurf

Identical pattern. Add to ~/.codeium/windsurf/mcp_config.json.

Get an API key first

curl -X POST https://api.ausdata.io/v1/register \
  -H "Content-Type: application/json" \
  -d '{"email":"you@example.com"}'

Free tier is 500 calls per month, no card.

What tools does the MCP server expose?

The ausdata MCP server exposes 5 core tools across all 9 Australian government data sources:

• search_datasets(query, source?) --- plain-English search
• describe_dataset(source, dataset_id) --- schema + valid filters
• get_data(source, dataset_id, filters, start_period?, end_period?) --- fetch records
• latest(source, dataset_id, filters) --- current snapshot
• list_curated(source) --- enumerate available datasets

Plus the composer endpoints (real_rate_regime, inflation_decomposition, etc.) as additional tools your agent can call.

A worked example

Once installed, you can ask Claude:

"What is the current real cash rate in Australia and how does it compare to a year ago?"

Claude will:

1. Call get_data("rba", "F1", {"series": "cash_rate_target"}, limit=1) to fetch the current cash rate
2. Call get_data("abs", "CPI", {"region": "australia", "category": "trimmed_mean"}, limit=1) for the trimmed-mean CPI
3. Compute the difference
4. Repeat for a year ago using start_period
5. Return the comparison

No hallucination. No "I do not have access to current data." The model has live ground truth.

Why this matters for production agents

Three failure modes that MCP grounding eliminates:

1. Stale data. Models trained more than 6 months ago will quote an outdated cash rate.
2. Hallucinated authority. Models sometimes invent plausible-sounding statistics when grounded data is unavailable.
3. Inconsistent answers. Without grounding, two calls to the same agent may return different statistics.

With ausdata MCP installed, all three disappear. The model becomes a wrapper around verified live data.

What this costs

The MCP server is free (open source on PyPI as ausdata-mcp). The API behind it has a free tier of 500 calls per month. Paid tiers start at $29/mo (10,000 calls) and $99/mo (100,000 calls).

A typical AI agent doing 50-100 grounded queries per day stays well within the free tier for personal use. Production agents handling customer traffic typically need the Embed tier at $99/mo.

Common gotchas

1. AUSDATA_API_KEY environment variable. If you do not set this, the MCP server starts but every tool call returns a 401. Easy fix; just set the env var.
2. Restart required. After editing the MCP config, restart your host (Claude Desktop, Cursor) so it re-reads the config.
3. Rate limits. The free tier caps at 10 requests per second. Most agent workflows hit this only during rapid back-and-forth.
4. Cold first hit on uncommon datasets. First call to an obscure dataset may take 5-15 seconds while our cache warms. Subsequent calls are sub-second.

Related resources

• /docs/mcp --- Full MCP server reference
• /cookbook/02-real-rate --- Worked example using the composer
• /sdk/python --- Python SDK if you prefer non-MCP integration

Honest read

If your AI agent answers questions about Australian macro, financial, energy, health, or regulatory data, you need grounded data. MCP is the cleanest way to provide it as of 2026. The ausdata MCP server is the canonical implementation for Australia.

Get a free API key → /register