How to Query Data Commons with MCP in Antigravity

Public statistics are useful only if an agent can find the right series and cite the source. Google’s Data Commons team now ships a hosted Model Context Protocol server so agents can search indicators and pull observations instead of guessing numbers.

This guide shows how to connect that server in Google Antigravity, confirm the tools loaded, and ask questions that match the tools the server actually exposes. The steps come from the Data Commons MCP docs, last updated September 22, 2026.

If you want the policy context for the UN System Data Commons launch, read how UN statistics became AI-ready. This article stays on the developer setup.

What the hosted server can do

The managed endpoint is https://api.datacommons.org/mcp. By default it returns data from datacommons.org, the public knowledge graph. A custom Data Commons instance needs a self-hosted server; the hosted URL will not point at your private graph.

The docs list six tools:

• search_indicators finds variables and topics for a place, such as health data in Egypt.
• search_child_indicators looks at contained places, such as census series for U.S. states.
• get_variable_metadata returns sources, dates, and other detail on a candidate indicator.
• get_observations fetches a time series for a variable and place.
• get_child_observations compares contained places, such as life expectancy across South American countries.
• get_multi_entity_observations covers directional links, such as exports from one place to another.

The server also exposes skills as MCP resources. Those skills are playbooks for single-place queries, sub-region queries, and multi-place flows. You do not need to paste Data Commons-specific instructions into the agent.

Unsupported today: non-geographical custom entities, events, free graph exploration, and data already formatted for charts. The docs also warn that agents can still make mistakes, so check the figures against the cited source.

Get a Data Commons API key

Every client in the official guide needs a free key for the api.datacommons.org domain.

1. Open apikeys.datacommons.org.
2. Request a key for the api.datacommons.org domain.
3. Store the key outside the repo. Do not commit it.

Antigravity reads the key from a header in your local MCP config. Treat that file like any other credential file.

Connect Antigravity to the hosted server

Install Antigravity from antigravity.google/download if it is not already on the machine. The docs cover both the IDE and the CLI.

Open ~/.gemini/config/mcp_config.json in the IDE or a text editor. Add the Data Commons server next to any other MCP servers you already use:

{
  "mcpServers": {
    "datacommons-mcp": {
      "serverUrl": "https://api.datacommons.org/mcp",
      "headers": {
        "X-API-Key": "YOUR_DATA_COMMONS_API_KEY"
      }
    }
  }
}

Replace the placeholder with your key. If the file already has an mcpServers object, add only the datacommons-mcp block so you do not wipe other servers.

Restart the IDE or CLI so it reloads the config. Then check discovery:

• /mcp tools should list the Data Commons tools.
• /mcp resources should list the skill resources.

If the tools do not appear, confirm the JSON is valid, the header name is X-API-Key, and the URL has no trailing path beyond /mcp.

Ask questions the tools can answer

The docs say the tools work best for comparisons and for exploring what data exists on a topic. Sample prompts from the same page:

• What health data do you have for Africa?
• What data do you have on water quality in Zimbabwe?
• Compare the life expectancy, economic inequality, and GDP growth for BRICS nations.
• Generate a concise report on income vs diabetes in US counties.

Start with a discovery question, then a fetch. A first prompt such as “What census data do you have for Canada?” lets the agent call search_indicators before it requests observations. A follow-up such as “List the population of Canada since 1964” maps to get_observations.

Ask for sources in the same thread. get_variable_metadata is the tool that returns source and date coverage. If the reply has a number and no source, ask which indicator and publisher it used.

Optional: run the sample ADK agent

Antigravity is the shortest path. The same docs also ship a sample agent based on the Agent Development Kit if you want a local web UI.

You need Git and the uv package manager. You do not install the ADK yourself. The run command downloads dependencies at start.

git clone https://github.com/datacommonsorg/agent-toolkit.git
cd agent-toolkit
uvx --from google-adk adk web ./packages/datacommons-mcp/examples/sample_agents/

Open the local URL printed in the terminal, often http://127.0.0.1:8000/, and type the same style of query. For a terminal-only session:

uvx --from google-adk adk run ./packages/datacommons-mcp/examples/sample_agents/basic_agent

To change the model, edit AGENT_MODEL in packages/datacommons-mcp/examples/sample_agents/basic_agent/agent.py. Behavior lives in AGENT_INSTRUCTIONS in instructions.py. Restart the agent after either edit. The docs say you still do not need custom Data Commons instructions, because the skills arrive as server resources.

Tips before you rely on a figure

Keep the API key out of screenshots and pull requests. A local mcp_config.json is enough for Antigravity.

Prefer hosted MCP for public datacommons.org queries. Switch to a self-hosted server only when you need a Custom Data Commons instance. The self-hosted options in the docs are a Python package, a Docker image, and the Custom Data Commons image on Cloud Run.

Do not expect chart-ready output. The current server does not return data formatted for graphic visualizations. Export the table and chart it in your own tool.

Skip event queries and custom non-geographic entities. Those are listed as unsupported.

Cross-check one observation against the publisher named in the metadata before you paste a number into a report. The Data Commons disclaimer is explicit: AI applications using the MCP server can make mistakes.

Conclusion

A free Data Commons API key, one block in ~/.gemini/config/mcp_config.json, and a restart are enough to point Antigravity at the hosted MCP server. From there, discovery tools and observation tools let an agent search public statistics and pull time series, while you still verify the source.

Use the sample ADK agent when you want a local UI or a model you can edit. Use a self-hosted server only when the public graph is not the dataset you need.

Sources

• Data Commons, “Query data interactively with an AI agent,” docs.datacommons.org/mcp, updated September 22, 2026.
• Data Commons, “Use MCP,” docs.datacommons.org/mcp/run_tools.html, updated September 22, 2026.
• Google, “Making global data easier to explore,” blog.google, September 17, 2026.
• Google for Developers, “Demo: DataGemma: Grounding LLMs with Data Commons data,” YouTube, October 18, 2024.