How Huwise’s MCP innovation underpins your agentic AI success

Since its introduction, the Model Context Protocol (MCP) standard has transformed how AI models and agents integrate with the wider data and IT stack. The rise of AI agents has further accelerated MCP’s importance, particularly around data marketplaces.

To find out more, we interviewed David Thoumas, Huwise’s CPTO and Co-founder. In this article he answers key questions about how the MCP server works, the benefits it delivers and how it underpins greater data consumption by AI agents and employees, more detailed understanding of data product usage, and simplified data marketplace administration.

“Model Context Protocol (MCP) is an open standard and framework, launched in late 2024, that standardizes the connection between AI models and data sources or tools. While originally created by Anthropic as an open source protocol, the Agentic AI Foundation is now responsible for its future development. It has been adopted by a range of major players in the AI industry.

Essentially, a MCP server provides a single connector that makes an organization’s data accessible to multiple AI assistants, without having to create individual integrations. That enables organizations to make data directly accessible in the tools that people are already using, as well as providing data seamlessly to AI agents. A good way to think of it as similar to a USB plug that works with a wide range of devices, from phones to laptops.”

“Huwise’s MCP server links a Huwise data marketplace to an organization’s tools through a common standard. This accelerates the discovery and use of its data by AI engines like Anthrophic’s Claude and Microsoft Copilot and enterprise tools like Microsoft Teams.

It enables these AI agents and LLMs to find, understand, and query a portal’s data products, while adhering to the organization’s governance rules. In practical terms agents can:

• Read the metadata of data assets: description, schema, and field definitions
• Search for datasets and data products in the catalog
• Query the data and extract results
• Conduct audits of the data marketplace and suggest areas for improvement

The MCP server actions any tasks requested by an AI agent on a client’s data marketplace. For example, it could quickly answer the query “Which datasets haven’t been updated in over six months?”

Currently, the Huwise MCP server powers:

• Our own agents:
  • Front end agents, such as the Huwy AI exploration agent, which accelerates time-to-insight for business users and data consumers on the data marketplace
  • Back-end agents that govern and administer the data marketplace and accelerate data product success: identifying future data products from usage signals, speeding up their creation, auditing and completing AI-ready metadata, and providing actionable recommendations
• Our clients’ own AI agents, created internally around their specific requirements.”
  

“Two main groups benefit from the MCP server – direct and indirect users, thanks to the data marketplace features it exposes.

Direct users include AI agents (such as Huwy, Claude, ChatGPT, internal agents) that connect to the MCP server to find and query data products on the marketplace, and developers and IT teams who benefit from having a standard access point, without the need to develop custom integrations.

Indirect beneficiaries include business users who can access data through the AI tool that they already use, without having to go through the data marketplace, and data teams, who benefit from wider consumption of their data products, without requiring specific, additional, development.

A good client example of the benefits is at the French Ministry of National Education, which uses the Huwise MCP server across several AI tools. It has also created the ability for its community to leverage data from their own environments and use it for educational purposes directly on its Huwise data marketplace.”

“Our MCP server was released in July 2025. It relies on our robust, well-documented APIs and makes them usable by AI agents. The APIs remain the foundation for standard technical integrations.

Naturally, we built the MCP using our own tools—namely, our APIs—which gave agents full access to the data right from the start. In practice, we found that agents tend to think in terms of tasks: for certain requests, they would chain together multiple calls. We saw this as a great opportunity to optimize the server in two ways. Firstly, reducing costs associated with the number of requests and secondly, speeding up data exchanges to improve the user experience.

So, after a year of monitoring usage patterns of our Huwy exploration agent, and our clients’ agents, we reorganized the tools this summer. Now, each one corresponds to a specific task and returns only the essentials. This means that the LLM just calls the right tools, resulting in faster and less costly interactions.

We have also enriched the context that the MCP server provides to agents. It now transmits aggregated information, such as filters and facets, which show the main values of a dataset and their distribution. This allows the agent to know what a dataset contains even before querying it, thereby improving performance. This rich context is available out of the box, as it is already generated by the native features of the data marketplace.”

“Three things set our MCP apart from others on the market. Firstly, it provides an advanced level of security, based on our own security standards in order to protect data. Secondly, it relies on our proven tools for leveraging data, such as APIs and technical documentation.

Thirdly, it provides access to governed data products, not just metadata, meaning the agent can query the data itself.

The Huwise data product marketplace is ideally suited to supporting AI agents. As data products are documented, governed, and maintained it delivers high-quality data to agents, all governed by built-in usage rules. Additionally, the marketplace understands real-world use cases (searches, questions, requests), which helps AI agents easily find the right data for their specific needs.”

“Security and governance are built-into the Huwise MCP server. As AI agents act on behalf of an identified human user, they are subject to that user’s permissions, meaning they can only view the same data on the marketplace as that person can. This applies whether the query is made directly on the marketplace or in a tool such as Microsoft Copilot via the MCP. The same governance rules apply. Access to data via the MCP is also tracked in the data marketplace to understand its impact as a usage channel.”

“We’re developing the Huwise MCP capabilities based on four priorities:

• Understanding which tools are used in a conversation. We want to know which tools the agent uses, depending on the request
• Providing the best possible context to AI agents. This applies to both the content being transmitted and the way it is structured
• Optimizing costs related to token consumption. This requires context that is as relevant and lightweight as possible. If an LLM needs to work on a task, it is only provided with the necessary tools for that task
• Limiting the number of tools available in a conversation. The larger the set of tools available to an agent during a conversation, the heavier the context, and the more expensive each request becomes

A good example of where we are focusing is our Automation API. Given its wide range of capabilities, our customers find this API really useful when automating administration of their data marketplaces.”

“The Living Context Layer is the knowledge that accumulates as the marketplace is used. It is then made available to agents to enhance their performance. This is what gives a Huwise data marketplace its lasting value.

The MCP server contributes to this in two ways:

1. It measures usage by agents. Every action an agent performs on Huwise passes through the MCP server, which allows it to be measured. For example, we can see that a data product is heavily used by agents while humans rarely access it – and that’s valuable information.
2. It makes the Living Context Layer accessible to agents. The MCP server shares some of this knowledge with agents, which they can then leverage directly.”

“We recently made the administration and configuration features of a data marketplace (i.e., the capabilities of our Automation API) available to agents through the deployment of our agentic AI framework. 

Moving forward we want to continue to expose and structure even deeper context for agents, while enabling users to generate data visualizations within our agents, and then make them shareable, and editable.

Overall, we’re committed to exploring how we can use the power of the MCP to meet our key objective of making access to reliable, understandable data simple and seamless for both humans, and AI agents. Watch this space for further updates!”