Published on: 12 sep 2025

Model Context Protocol at Blenddata: This is how we make AI talk to our data

AuteurKevin BeversFunctiePlatform Engineer

Tag(s)

Expert GenAI

In the world of artificial intelligence and data analytics, we are constantly looking for ways to make AI tools more effective. One of the biggest challenges is connecting AI systems to existing data sources and tools. The Model Context Protocol (MCP) offers an elegant solution for this. In this blog post, we take you through our experience implementing MCP at Blenddata and show how this protocol is revolutionising the way AI can work with data.

AuteurKevin BeversFunctiePlatform Engineer

Tag(s)

Expert GenAI

What is the Model Context Protocol (MCP)?

The Model Context Protocol (MCP) is a standardised way to give Large Language Models (LLMs) direct access to tools, APIs, databases, documentation and other data sources. Instead of having to copy and paste information, the AI can directly use these tools and sources to get context and perform tasks. This opens up a world of new possibilities.

A common problem at companies is that valuable information and tools are available, but not accessible to AI tools. Employees have to manually look up, copy and paste information to get AI input. This takes time and hampers productivity. MCP solves this by connecting your LLM directly to your existing tooling and data sources.

Why is MCP important, what opportunities does it offer?

MCP is important for three core reasons:

Standardisation – Before MCP, there was no standardised way to connect LLMs to tools and data sources. Each integration was a custom solution. MCP provides a universal standard supported by the entire industry.
Direct tool integration – MCP allows your LLM to communicate directly with your existing tooling, APIs and data sources. This means your AI can not only talk, but also act using your tools.
Enabling new workflows – These connections open up new use cases that were previously not possible, such as querying databases directly via natural language or AI assistants accessing up-to-date documentation and codebases.

Concrete examples:

Talk to your data – Ask questions to your database via natural language and get instant answers
Development support – AI accessing your API docs, dbt models and codebase for better project understanding

At Blenddata, we saw this as an opportunity to make our workflows more accessible and improve the development experience by integrating MCP into our daily processes.

How did we implement MCP at Blenddata?

During our research into LLMs working with our data, MCP emerged as the best solution. We wanted to explore how to make an LLM work with our data from data pipelines, and MCP offered exactly what we were looking for: a standardised way to connect AI tools with data sources.

Our implementation in 5 steps:

Step 1: Setting up data context with dbt MCP

Tool: Existing dbt MCP server (modified)
What we did: We automatically documented our Blenddata data stack through dbt docs of our data models and their interrelationships. To this documentation, we connected an MCP server, which now allows the LLM to query the context of the data models.
Customisation: We modified the existing server so that we can use a tool call to retrieve the project description page, which works much better for the AI.

Step 2: Database access via PostgreSQL MCP

Tool: Existing Microsoft MCP server for PostgreSQL (modified)
What we did: For our database running on PostgreSQL on Azure, we used an existing MCP server from Microsoft and modified it slightly.
Security measure: We disabled data modification and deletion tools for security.

Step 3: Selecting test data

Data: internal AFAS data on productivity, revenue and project hours
Reason: As we already had dashboards for this, this was the ideal way to test “chatting with your data”.

Step 4: Configure AI workflow

Tools: System prompts and fine-tuned prompts
What we did: We finetuned prompts and deployed workflows using system prompts (rules). The AI first retrieves dbt context information based on the metrics being queried, and then creates an SQL query with it.

Step 5: Model comparison and client development

Tools: Different LLM models (o3-mini, GPT-4.1, GPT-4.1-mini) and proprietary Python UI
What we did: We built our own MCP client – a Python UI using MCP servers, and tested different models to find the best performance-to-cost ratio.

What did we learn from our MCP implementation?

Our test results show that MCP can be extremely useful in development scenarios, but additional steps are needed for dashboard end products.

Our 4 key insights:

MCP is already a game-changer for development – We recommend every company to connect their AI assistant to existing tooling. For instance, I have my Notion, Terraform documentation and dbt docs attached to my Cursor IDE, which allows me to easily query information without pasting back and forth.
GPT-4.1-mini offers the best performance-to-cost ratio – During our model comparison, GPT-4.1-mini proved to be the best choice for MCP implementations where cost is a factor, while tool usage accuracy was identical to GPT-4.1.
Hybrid approach works for dashboards – We can use MCP as a dashboard chatbot where all the data is already pre-populated via SQL queries from analysts. The LLM gets these predefined queries as tools, after which end users can chat with them.
Reliability remains the biggest challenge – MCP offers valuable internal applications, but is not yet ready for zero-shot customer-facing applications. Reliability and accuracy remain the biggest challenges, although this can change quickly with the continuous development of LLMs.

TLDR

✅ Works well:

Developer tools and code comprehension
Interaction with documentation and APIs
Bridge to existing tooling (CLIs, documentation)

⚠️ Challenges:

Text-to-SQL is not always 100% reliable
Context inconsistency between different tools
Hallucination risk in complex queries

🔒 S ecurity guidelines:

Read-only access to databases
Test with small datasets first
Log which tools are used

How can you start using MCP servers?

Want to get started with MCP yourself? We’ll show you how to set up your own MCP setup in just a few steps. We use Notion as an example, but you can apply this process to other MCP servers.

Step 1: Download Cursor IDE Download Cursor IDE fromcursor.com/downloads. This is our MCP client that we will use.

Step 2: Choose an MCP Server For this tutorial, we will use the Notion MCP server from makenotion/notion-mcp-server. This is a good choice because Notion is accessible to most people.

Step 3: Set up Docker Desktop Make sure Docker Desktop is up to date. We are going to use the Docker MCP toolkit (beta) to simplify installation. You can also do this via the mcp config .json that most clients have.

Step 4: Add the Notion MCP server In Docker Desktop, go to the MCP toolkit and add the Notion MCP server. This process is visual and user-friendly.

Step 5: Configure the Notion MCP server You will need a Notion API token. Go to notion.so/profile/integrations to create it. Follow the instructions on the Notion MCP GitHub page for the exact steps.

Step 6: Connect Cursor to the MCP toolkit In the MCP toolkit, go to the clients tab and click “connect” at Cursor. This adds the docker mcp toolkit to Cursor’s mcp.json.

Step 7: Restart Cursor (optional) Restart Cursor to make sure everything is loaded correctly.

Step 8: Check the connection Go to File > Preferences > Cursor Settings > Tools & Integrations to see if the MCP server is connected correctly.

Step 9: Test it out! Start a new chat (Ctrl + I) and try it out. You can now ask questions about your Notion data and the AI has direct access to your Notion workspace.

Sample prompts to try:

“What are my latest tasks in Notion?”
“Show me the most important notes from this week”
“Help me create a new page structure for my project”

Our recommendations for beginners

Start small: Start with simple mcp servers and use cases
Test thoroughly: Validate results returned from tooling (it’s still an LLM!)
Focus on development: Use MCP as a developer aid, not as an end-user tool

Tip: Using MCP servers to give context to your AI assistant is ideal for new team members to quickly understand complex projects. For example, “I’m going to work on project X, help me understand the data models.”

We make your data accessible to AI

Blenddata is a data engineering consultancy that specialises in building reliable data pipelines and modern data architectures. We help companies make their data accessible and usable.

With our experience in MCP, we can help your organisation start using AI tools. Whether it’s dbt documentation, databases or other sources, we can help you harness the power of AI in your day-to-day work.