Keboola MCP Server: Best Practices and Frequently Asked Questions

‍After 10 days since launching the Keboola MCP (Model Context Protocol) Server, we've gathered the most common questions from our data community. This article combines practical answers with best practices inspired by successful AI-assisted development patterns, helping you get the most out of your AI-powered data workflows.

Introduction

The Keboola MCP Server represents a new paradigm in data operations - enabling AI agents to directly interact with your data platform while maintaining security and governance. As early adopters have explored its capabilities, certain patterns and questions have emerged consistently.

This guide addresses the top questions we've received and provides actionable best practices to accelerate your success with AI-assisted data operations.

Top 5 Questions from the First 10 Days

1. How much guidance do I need to provide the AI agent when working with data using the MCP Server?

The amount of guidance depends on your task complexity:

For basic operations (querying tables, simple transformations):

• Minimal guidance needed - ideally just specify the table names and desired outcomes in the business logic. For example most AI tools know how to create market segmentation and other outcomes without any special commands, as far as it understands the data set it is using.
• Example: "Show me total sales by month from the orders table"

For complex workflows: More detailed instructions help, including:

• Specific column names and data types
• Business logic and transformation rules
• Expected output format
• Performance considerations for large datasets

Important Note: AI models are non-deterministic by nature. While the numbers and calculations will usually be correct (based on the metrics used), the presentation format may vary between runs unless explicitly specified. For example, the same query might return results with different column orders, date formats, or decimal places unless you define these requirements.

Example of format variation:

To ensure consistency, specify your format requirements:

The MCP server handles the technical implementation details, so focus on describing what you want rather than how to do it.

Best Practice: Start with minimal guidance for simple tasks, then add context as complexity increases. Always specify format requirements for production workflows to ensure consistent outputs.

2. Can I plug it into multiple projects?

No, currently Keboola only supports connection to a single project. However, you can always switch between the projects. Some AI agents support multiple active MCP servers in parallel, but we do not recommend this as the performance and response rate may result in a poor experience. 

3. What if my AI breaks something?

Keboola has multiple layers of protection to safeguard your data:

Built-in Keboola safeguards:

• Version control: Your configurations and transformations in Keboola are versioned and every change is logged.
• Rollback capabilities: You can revert to previous working states
• Workspace environments: You can always test changes in isolation just within the AI agent before applying to production
• Access controls: MCP operations respect your Keboola permissions
• Audit logs: Track what changes were made and when

AI tool-level permissions: Most AI tools (like Cursor and Claude) provide an additional safety layer through tool permissions:

• Permission prompts: By default, the AI agent will ask for your approval before executing any action using a particular tool that reads or modifies data or configurations
• Granular control: You can approve or deny each specific action
• Safe exploration: You can let the AI read and analyze freely while requiring approval for any write operations

Example permission flow:

This means you can always take the safe approach - if you're unsure about an operation, simply refuse the permission request. The AI will then suggest alternatives or ask for clarification.

Best Practice:

• Keep permissions for “retrieve”, “create” and “get” tools set to “Allow always” and keep  the “update”, “set” and “run” tools asking you a permission every single time until you get confident working with the MCP server. 
• Create a simple rollback procedure documentation that your team can follow if needed - this can be simply automated using the Keboola MCP server as well.

4. What is the optimal way to build via MCP?

For efficient development, follow these principles:

Start small and iterate

• Build simple queries first
• Test each component before combining
• Use descriptive naming conventions

Leverage templates

• Create reusable transformation patterns
• Document common workflows
• Build a library of tested solutions

Structure your requests clearly

Use consistent patterns

• Standardize table and column naming
• Follow data modeling best practices
• Document business logic

5. Is there anything like saved searches so I don't need to repeat my commands?

While MCP is just a set of tools having no "saved searches" per se, you can:

1. Create Flows in Keboola: Save complex workflows as Keboola Flows that can be triggered via MCP at any time or that can be run automatically based on a schedule.
2. Use transformation templates: Build reusable SQL/Python transformations
3. Maintain a command library: Keep a personal collection of frequently used MCP commands and save them for example in AI rules in Cursor. 
4. Leverage flows: Group related transformations into flows that can be executed as a unit.
5. Use some of the Keboola pre-defined prompts: If you use Claude, in the toolbar below the chatbox, you can find a few fine-tuned and well tested prompts Keboola as created for you for the situations like when you need to run a Data Quality Assessment in your Keboola project or when you need to perform a health check. Our prompts ensure you get precise and comprehensive information in a consistent format. 

Pro tip: Create a "cookbook" of common data operations in your organization that can be easily referenced and adapted for new use cases. For instance you can integrate the library built within a third-party knowledgebase supporting MCP Server integration as well, such as Confluence. 

Advanced Best Practices: Learning from AI Development Tools

Based on successful patterns from tools like Cursor AI, here are advanced techniques for maximizing your MCP Server effectiveness:

1. Break Down Complex Tasks into Atomic Operations

Instead of:

Breakt it down:

2. Provide Business Context Explicitly

Less effective:

More effective:

3. Check and Validate at Each Step

Implement a verification workflow:

4. Explain Domain-Specific Logic

5. Provide Sample Expected Output

6. Document Edge Cases

7. Use Iterative Refinement

8. Maintain Context Between Sessions

Bonus Question: Which AI model is the best with the Keboola MCP server?

The landscape of AI models is evolving at such a rapid pace that we intentionally avoid making direct comparisons between models. What's cutting-edge today might be surpassed tomorrow, and model capabilities can vary significantly based on updates and specific use cases.

What we've observed from early adopters:

Different models excel in different scenarios. For most tasks, given the higher context window, and code generation abilities, we recommend Claude. Auto-mode in Cursor is just fine… 

• Some models handle complex SQL generation and data relationships particularly well
• Others shine when working with Python transformations or debugging code
• Certain models offer faster response times for simple queries
• Some provide more thorough analysis for complex data investigations

The Personal Factor:

Interestingly, the "best" model often depends heavily on personal experience and prompting style. We've noticed that:

• Users who write detailed, structured prompts might get better results with one model
• Those who prefer conversational, iterative approaches might excel with another
• Some models respond better to technical jargon, while others prefer plain language
• Your background (data engineer vs. business analyst) can influence which model feels more intuitive

Each person develops their own "prompting personality" over time, and certain models align better with different communication styles.

Our recommendation: Rather than declaring a "best" model, we encourage you to:

1. Start with a model you're already familiar with
2. Test it with your actual use cases and natural prompting style
3. Measure performance on your specific data patterns
4. Be prepared to switch as new models emerge
5. Share prompting techniques with your team—what works for one person might help others

‍

Conclusion

The Keboola MCP Server opens new possibilities for AI-assisted data operations. By following these best practices and learning from early adopters, you can accelerate your journey toward more efficient, intelligent data workflows.

Remember: success comes from clear communication with your AI agent, iterative development, and leveraging Keboola's built-in safety features. Start simple, document your patterns, and gradually build toward more complex operations.

Important reminder: The principle of "garbage in, garbage out" still applies—even with AI. The quality of your results depends on the quality of your data, the clarity of your instructions, and the context you provide. AI amplifies what you give it, so invest time in proper data preparation and thoughtful prompting.

Ready to get started? Try implementing one of these best practices in your next data project. We'd love to hear about your experience!

Have recommendations or questions we didn't cover? Share your insights with us at talkto@keboola.com. Your feedback helps shape our best practices and benefits the entire community.

‍