Google’s MCP Server aims to cut AI hallucinations with verifiable data

Google has revealed a Model Context Protocol (MCP) Server for its Data Commons platform. The MCP server allows developers and AI systems to access publicly available datasets using natural language. 

The Data Commons platform houses census figures, climate datasets, and statistics from bodies such as the UN. The tool aims to make integrating high-quality data into AI training pipelines and applications easier.

Google’s MCP Server to reduce AI hallucinations

The Data Commons platform was introduced in 2018 to consolidate public datasets from international organizations, government sources, and local administrations. Accessing the datasets has been a challenge that requires extra technical expertise. However, the release of the MCP server has made it easier for developers and AI agents to query those datasets using natural language, expanding its accessibility to AI systems and users.  

Wish you could easily find answers to your most complex data questions or generate marketing plans from your data? Our new MCP server connects your GA data directly to LLMs like Gemini, making it possible.

In just 8 minutes, @matt_landers guides you through MCP server setup:… pic.twitter.com/sw9vDM5Vbs

— Google Analytics (@googleanalytics) September 19, 2025

Until today, AI systems have been trained on unverified web content, which increases the risk described as hallucination when combined with the tendency to fill in blanks to generate outputs, according to a study by Cornell University.

Google has confirmed that the release of MCP Server has been designed to provide reliable and verifiable datasets that can feed AI systems with real-world information. Training pipelines and other practical applications will now be able to use publicly available data, such as census figures and climate statistics, in an easier and verifiable manner. 

“The Model Context Protocol is letting us use the intelligence of the large language model to pick the right data at the right time, without having to understand how we model data and how our API works.”

-Prem Ramaswami, Head of Google Data Commons

The Model Context Protocol (MCP) tool debuted in November 2024 by Anthropic as an open-source platform to provide a framework for AI systems to access structured data from multiple sources, including content repositories, business tools, and application environments. Since its launch, major AI startups, including OpenAI, Microsoft, and Google, have adopted the tool to connect AI models to external data sources. 

Unlike other companies that tested the MCP tool directly with their models, Google’s Data Commons team integrated the standard directly into its platform, resulting in a dedicated MCP Server now open to developers. 

MCP Server is now available to developers

Google has also partnered with the ONE Campaign, a non-profit organization focusing on public health and economic opportunities in Africa, to build the One Data Agent tool.

Using the Data Commons MCP Server, the tool harnesses multiple datasets from health and finance in plain language. Ramaswami revealed that the collaboration helped accelerate the development process since the non-profit had experimented with MCP on its own server before approaching Google. 

The search engine company confirmed that the MCP Server is compatible with any large language model (LLM) and gave developers the go-ahead to experiment using a sample agent in Google’s Agent Development Kit (ADK), available on Colab Notebook. The firm added that the MCP Serve can be accessed via the Gemini command line interface or any MCP-compatible client using the PyPl package.

Google’s latest launch would help reduce reliance on unverified online data, improve the reliability of AI systems, and allow researchers and organizations to incorporate trusted data more easily. The MCP Server has been made available for developers worldwide as open source for trials.

OpenAI also published its own guide on building MCP servers for ChatGPT and API integrations, enabling developers to extend AI models with new data sources.

The guide demonstrated ways to create a remote MCP server using Python and FastMCP, with examples of integrations into the OpenAI chatbot. OpenAI also highlighted the risks of using custom MCP servers, including malicious servers prompting injections to steal sensitive information, and urged developers to exercise extra caution or connect with trusted or official servers only.

If you're reading this, you’re already ahead. Stay there with our newsletter.