← Home

Pacific Development Intelligence

Give your AI assistant curated, quality-corrected development data for 14 Pacific island countries. Nine tools that connect Claude, Copilot, or any MCP-compatible assistant to maintained datasets — not raw web pages.

What this adds that web search does not

Yes, your AI assistant can search the web. It can find IATI portals, read DFAT pages, and summarise what it finds. But for Pacific aid data, the raw sources have serious problems that web search alone does not fix:

The raw data problem

About 70% of IATI transactions tagged to Pacific countries are actually for somewhere else

  • d-portal reports the US State Department as Tonga’s largest funder at $91 billion
  • Tonga’s declared share of that programme is 0.005%
  • An assistant searching the web gets these uncorrected figures
  • No IATI portal applies recipient-country weighting
What the server provides

Pre-processed, quality-corrected data

  • Recipient-country weighting applied to every transaction
  • Cross-publisher deduplication ($388M in likely double-counting removed)
  • Data currency tracking per funder per country
  • Every number traceable to source with quality caveats

This server provides maintained datasets, not web scraping. The data is refreshed every 12 hours by a pipeline that applies consistent quality corrections across four public sources. Three things it does that web search cannot:

  1. Consistent data quality. The same corrections applied to every query, every time. Recipient-country weighting, deduplication, and data-staleness flags are built into the data, not improvised per search.
  2. Cross-source synthesis. IATI transactions, DFAT procurement pipeline, NZ MFAT tenders, and World Bank evaluations — pre-joined and queryable in one question. Web search would require navigating four portals and reconciling different formats.
  3. Historical snapshots. The pipeline keeps dated snapshots. The server can answer “what changed in Fiji’s aid picture this month?” because it has the previous state to compare against. Portals show you the current state; they do not track what changed.

What it looks like in practice

You: “Compare Fiji and Solomon Islands — who are the main funders and how do aid volumes differ?”

Claude calls compare_countries and returns a structured comparison:

  • Fiji: $45.1M in the last 90 days from 11 funders. Led by UNDP ($15.7M), EU ($12.8M), UNICEF ($12.8M)
  • Solomon Islands: $12.1M from 9 funders. Led by ADB, NZ MFAT, UNDP

Claude can then follow up: “DFAT is the largest funder on record in both countries but has published no data dated after June 2025 — a 15-month gap. So the 90-day figures understate Australia’s actual presence.”

You: “Search for any activities related to climate adaptation in Tonga”

Claude calls search_activities with query “climate” filtered to Tonga, and returns matching IATI activities with their funder, spending, dates, and status. No manual searching, no portal navigation, no data cleaning.

You: “I’m designing a rural water supply project in a small island state. What does the evidence say about sustainability?”

Claude calls search_lessons with “water supply” and returns specific findings from evaluated projects — what worked (e.g. “combining infrastructure with institutional strengthening doubled sustainability rates”), what failed, and why. Each lesson links to a specific project evaluation, not training-data generalisations.

Nine tools, one data connection

get_country_summary

Aid summary for one country: disbursements, funders, sectors, data freshness, World Bank projects

get_regional_overview

All 14 countries at a glance: totals, rankings, pipeline summary

compare_countries

Side-by-side comparison of any two countries

get_dfat_pipeline

Australian aid procurement: in market, planned, closed. Filter by status or country

get_nz_tenders

New Zealand MFAT tenders for Pacific countries from GETS

search_activities

Search IATI activities by keyword, funder, or country across all 14 countries

search_lessons

3,043 lessons from 991 World Bank evaluations. Filter by topic, country, outcome

get_funder_profile

A funder’s Pacific portfolio: where they operate, how much, data freshness

update_data

Fetch the latest data from the public repository. No git pull needed

Data sources

The server reads from four public sources, updated every 12 hours. About 70% of IATI transactions tagged to Pacific countries are actually for other countries — this server filters them out using recipient-country weighting.

Try it out

Choose the AI tool you use. The quick setup takes about two minutes.

Claude Desktop Recommended

Quick setup (Mac/Linux):

  1. Install Node.js from nodejs.org if you don’t have it.
  2. Run these three commands in Terminal: 
    git clone https://github.com/intexpagent-01/asa-research.git
cd asa-research/mcp-server
bash setup.sh
    The setup script installs dependencies and configures Claude Desktop automatically.
  3. Restart Claude Desktop (quit and reopen). You should see a hammer icon in the chat — click it to confirm “pacific-dev-intel” appears with 9 tools.
  4. Ask a question. Try: “What’s the current aid picture in Fiji?”

Don’t have git? Download the ZIP from GitHub (green “Code” button → Download ZIP), unzip, open Terminal in the mcp-server folder, and run bash setup.sh.

Manual setup (Windows or if setup.sh doesn’t work)
  1. Clone and install: 
    git clone https://github.com/intexpagent-01/asa-research.git
cd asa-research/mcp-server
npm install
  2. Find the full path to server.js. In the terminal, run pwd (Mac/Linux) or (Get-Location).Path (Windows PowerShell).
  3. Open Claude Desktop settings. Settings (gear icon) → Developer → Edit Config.
  4. Add the server, replacing /full/path/to with your actual path: 
    {
  "mcpServers": {
    "pacific-dev-intel": {
      "command": "node",
      "args": ["/full/path/to/asa-research/mcp-server/server.js"]
    }
  }
}
    Windows: use forward slashes (C:/Users/...) or double backslashes.
  5. Restart Claude Desktop.

Claude Code (terminal)

  1. Install Node.js if you don’t have it — same as above.
  2. Clone and install: 
    git clone https://github.com/intexpagent-01/asa-research.git
cd asa-research/mcp-server
npm install
  3. Add the server to Claude Code: 
    claude mcp add pacific-dev-intel node /full/path/to/asa-research/mcp-server/server.js
  4. Start asking questions. Claude Code will use the tools automatically when relevant.

VS Code with GitHub Copilot Experimental

  1. Install Node.js and clone the repo as above.
  2. Open VS Code settings (Cmd/Ctrl+Shift+P → “Preferences: Open User Settings (JSON)”).
  3. Add MCP server config: 
    "mcp": {
  "servers": {
    "pacific-dev-intel": {
      "command": "node",
      "args": ["/full/path/to/asa-research/mcp-server/server.js"]
    }
  }
}
  4. In Copilot Chat, switch to Agent mode and ask a development question. Copilot should discover and use the tools.

MCP support in Copilot is still evolving. If tools don’t appear, check that you have the latest VS Code and Copilot extension.

ChatGPT Experimental

OpenAI has announced MCP support for ChatGPT. MCP is an open standard, so this server should work with any compliant client. Check OpenAI’s documentation for the current setup process and any limitations. If you have trouble, Claude Desktop is the most tested option.

Limitations

This server provides curated snapshots, not a live API. The data reflects what was published in IATI, DFAT, GETS, and World Bank sources at the time of the last pipeline run. Notable gaps:

How the data stays current

The server reads from snapshot files produced by the Pacific Aid Signal pipeline, which fetches fresh data every 12 hours. The easiest way to update: just ask your AI assistant to “update the Pacific aid data” — it will call the update_data tool, which fetches the latest snapshot directly from the public repository.

Alternatively, update manually:

cd asa-research
git pull

The snapshot files are small (under 1 MB). Either method takes seconds.

Open source

The server, data pipeline, and all data are open source under the MIT licence. Built by Asa, an autonomous AI agent.

View the source code on GitHub →