Aus Data Agent MCP
J-King-Dottie/aus-data-agent-mcp
Acesso estruturado a dados públicos australianos e macro globais para análise de IA
Ver no GitHubServidor MCP unificado que integra múltiplas fontes de dados públicos australianos (ABS, RBA, DCCEEW) e globais (OECD, Banco Mundial, FMI, Comtrade) para agentes de IA. Oferece fluxo estruturado de descoberta, validação e armazenamento de variáveis econômicas e estatísticas com rastreabilidade completa.
Instalação
npx convez-ai J-King-Dottie/aus-data-agent-mcp1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Análise econômica e estatística baseada em evidências para agentes de IA
- Recuperação e validação de variáveis macroeconomômicas de múltiplas fontes
- Pesquisa comparativa internacional entre dados OECD, Banco Mundial e fontes locais
Veredito

Ferramenta especializada e bem-arquitetada para análise de dados com rigor metodológico. Excelente para contextos de pesquisa econômica e agentes que precisam de validação serial de variáveis, mas requer compreensão técnica e domínio de estatística pública.
README
A tradução para o português deste README ainda está na fila.
Aus Data Agent MCP
A unified MCP server for AI agents working with Australian public data, plus global macro sources for context and comparison.
Produced by Dottie AI Studio.
Built on existing open source work including seansoreilly/mcp-server-abs and hanlulong/openecon-data.
The project is built for evidence-led variable validation:
user asks for a metric -> search one candidate dataset -> inspect metadata
-> retrieve one source slice -> inspect one raw artifact -> narrow to one exact variable
-> show validation preview -> user approves -> save replayable variable recipe
The MCP does data discovery and retrieval. Agent behaviour lives in AGENT_SYSTEM_PROMPT.md, and development guardrails live in AGENTS.md.
For AI Agents
Start with these files:
- AGENT_SYSTEM_PROMPT.md: the agent system prompt for careful economic/statistical analysis.
- AGENTS.md: project architecture and development rules.
- backend/app/unified_mcp_server.py: MCP tool surface and tool descriptions.
- .mcp.json: project-scoped MCP config.
- UNIFIED_CATALOG_FULL.json: checked-in unified catalog snapshot.
Core tool flow:
one variable request -> search_catalog -> choose one dataset -> get_metadata
-> retrieve -> inspect_artifact -> narrow_artifact
-> validation summary -> save_validated_variable
Search results are candidate pools. The AI analyst chooses one dataset for one variable, inspects structure, retrieves real source data, narrows it to the minimum exact slice, and asks the user to approve before saving.
The MCP tools are intentionally single-call and serial:
- no batch catalog searches
- no batch metadata inspection
- no batch retrieval
- no batch artifact inspection
- no batch or parallel artifact narrowing
If a user asks for two columns, such as dwelling completions for houses and apartments, the app should validate them as two variables. Each variable gets its own exact narrowed recipe. After approval, later analysis can replay both recipes and join them into one table or spreadsheet.
Runtime artifacts are temporary working files. Validated variables are durable recipes containing the source, dataset, exact retrieval arguments, exact narrowing filters, transformation logic, variable type, and a short recreation summary. Future runs should replay recipes against source data rather than reusing broad old artifacts or discovery transcripts.
Data Coverage
The checked-in unified catalog currently includes 31,309 entries:
| Provider | Catalog entries | Typical use |
|---|---|---|
| World Bank | 28,413 | Global macro, development, population, poverty, education, health, trade, energy, climate, finance, governance |
| OECD | 1,470 | International comparisons, regions, labour, education, health, productivity, government, transport, income, prices |
| ABS | 1,221 | Australian official statistics: labour, population, CPI, national accounts, business, trade, housing, census, health, agriculture, industry |
| IMF | 132 | Macro-financial context: debt, money, credit, trade categories, AI preparedness, financial openness |
| Reserve Bank of Australia | 71 | Rates, exchange rates, money, credit, bank balance sheets, payments, financial markets |
| DCCEEW | 1 | Australian Energy Statistics Table O: electricity generation by fuel type |
| UN Comtrade | 1 | Goods imports/exports by partner and HS commodity code |
Useful discovery keywords:
ABS, Australian Bureau of Statistics, labour force, employment, unemployment, CPI, inflation, wages, population, migration, births, deaths, national accounts, GDP, household spending, retail, building approvals, housing, lending, business indicators, international trade, exports, imports, census, agriculture, energy generation, electricity by fuel, RBA, cash rate, exchange rates, interest rates, credit, financial institutions, OECD, World Bank, WDI, IMF, Comtrade, HS code, partner country.
Architecture
backend/app/unified_mcp_server.py: unified MCP server.backend/app/unified_catalog.py: catalog loading and SQLite FTS search.backend/app/domestic_data.py: ABS and Australian domestic retrieval.backend/app/macro_data.py: OECD, World Bank, IMF, RBA, UN Comtrade, and related macro retrieval.scripts/build_unified_catalog.py: rebuilds the unified catalog and FTS index.frontend/: app layer being developed on top of the MCP.
The checked-in catalog is a discovery index, not a raw data mirror. Retrieved results are runtime artifacts used only to inspect and narrow one variable. Raw retrieved artifacts are inspect-only; only narrowed artifacts may be used for calculation or validation previews.
Quick Start
python3 -m venv .venv-wsl
source .venv-wsl/bin/activate
python3 -m pip install -r backend/requirements.txt
Run the MCP server:
python -m backend.app.unified_mcp_server
Example project MCP config:
{
"mcpServers": {
"ausdata": {
"command": "python",
"args": ["-m", "backend.app.unified_mcp_server"],
"cwd": "."
}
}
}
Refreshing Catalog Assets
Refresh the unified catalog and FTS index:
python3 scripts/build_unified_catalog.py
Refresh the local UN Comtrade metadata bundle:
python3 scripts/build_comtrade_metadata.py
App Layer
The repo still includes the app code in frontend/ and the supporting backend code in backend/. The app is being developed as an interface over the MCP/retrieval stack, but the public-ready surface for agents is currently the MCP plus the agent system prompt.
The web app is evolving into an AI-assisted economic modelling workspace:
-
Left pane: modelling projects and project switching.
-
Centre pane: AI workspace for question definition, discovery, variable construction, model execution, and results.
-
Right pane: compact model builder showing validated variables, assumptions, and mathematical links.
-
Supabase schema: supabase_modelling_workspace.sql creates projects, chat history, validated variables, assumptions, model graph, runs, run joins, and export records.
-
Frontend dev server:
http://127.0.0.1:3000 -
Backend:
http://127.0.0.1:5000
Daily local development from PowerShell with WSL:
wsl bash -lc "cd /home/projects/abs-mcp && ./scripts/dev-backend-wsl.sh"
wsl bash -lc "cd /home/projects/abs-mcp && ./scripts/dev-frontend-wsl.sh"
To open two visible terminal windows, one for the frontend and one for the backend, both with auto-reload:
powershell -ExecutionPolicy Bypass -File .\start-visible-dev.ps1 -SkipInstall
Smoke Checks
python3 -m py_compile backend/app/*.py backend/app/storage/*.py scripts/build_unified_catalog.py
npm --prefix frontend run build