Indeed Shares Keys to Building Effective MCP Apps

Dustin Mihalik, an AI platform engineer at Indeed, works on ensuring guardrails, gateways, and compliance for AI initiatives at the world's leading job site.

His team develops MCP applications for various AI models, including Claude, ChatGPT, and their internal agent named 'Career Scout', specifically to bridge the gap between AI chat interfaces and structured user experiences for job seekers.

To prevent a poor user experience, Indeed evaluated how models handle job details after finding that Claude and similar interfaces often omit branding and direct links to external sites.

Models tend to keep users within their own environment, making it difficult to integrate external site links consistently, necessitating extensive evaluation efforts from Indeed to ensure models reliably link to job details.

Utilizing MCP apps and SDKs gives developers the ability to control specific interface elements, such as an 'Apply' button, enhancing the user journey.

Developers can prominently display important information at the top of the interface, and modal windows can enable users to view job details without exiting the chat, significantly improving the overall user experience.

A common pitfall is to directly call existing APIs for data, which prevents the AI model from understanding the information displayed to the user.

If the model is unaware of the data being shown, it cannot effectively answer follow-up questions, making the first rule of building MCP apps to provide any user-displayed data to the model as structured content.

Developers must use the MCP specification to return both structured content and resource URIs pointing to HTML, ensuring both the model and the interface remain synchronized.

This dual return mechanism prevents the model from generating incorrect or misleading information by maintaining consistency between API changes and the data output it receives.

By specifying in the tool description that results are automatically rendered as UI components, developers prevent output clutter.

Explicitly stating that 'results are automatically displayed as UI components' in the tool description helps prevent the model from attempting to render UI components itself, proactively avoiding conflicts between model output and application interface.

The AI model typically lacks visibility into specific user actions, such as clicking a 'View Details' button or initiating a job application.

To address this, the 'update model context' method enables the transmission of string-based events, allowing the model to track user interactions.

For example, in a shopping cart application, tracking items and total costs through this method empowers the model to accurately answer follow-up questions about the user's cart.

Users frequently assign complex tasks to AI models, such as searching across multiple cities and applying various industry filters.

While text-based MCPs excel at multi-step tasks by generating comprehensive tables at the end, coupling results directly to a UI can cause the model to prematurely stop after a single action, perceiving the task as complete.

Therefore, to maintain a fluid agentic workflow, it is crucial to avoid multiple UI carousels that might interrupt the model's multi-step processing capabilities.

Following OpenAI Apps SDK guidance, developers should separate data processing from UI rendering in MCP applications.

This approach involves the model performing data operations using text-based tools first, with rendering tools then triggered separately to display the processed data, thus reducing complexity by allowing the model to filter data before selecting what to render.

In practice, this means creating distinct tools such as a 'search jobs' tool to fetch raw data without generating UI, and a separate 'render jobs' tool that accepts specific IDs to display results.

This design allows the model to fetch numerous results and then pass a filtered subset to the renderer, necessitating updated tool descriptions to ensure the correct calling order.

Qualitative explanations for selections can be integrated into the interface by designing rendering tools to accept additional fields such as reasoning or highlighted sections.

This capability enhances the user experience by injecting a unique character into the interface, making it applicable across various industries like e-commerce for personalized recommendations.

A primary takeaway for building MCP apps is to prioritize identifying data needs before planning the user interface, treating UI rendering as a side effect or result of the model's data exploration.

This approach advocates for using small, composable tools rather than large, overloaded ones, coupled with simple tool descriptions to prevent model confusion while maintaining flexibility.

When developing MCP apps, the focus should be on data before UI, a concept that may seem counterintuitive to traditional development, but ensures the model has a clear understanding of what it is interacting with.

Employing small, composable tools, such as multiple search tools and one or two render tools, allows for simpler descriptions, preventing the model from becoming overloaded and enhancing overall clarity and efficiency.

Minimizing the model's processing workload is achieved by building specialized, lightweight tools for singular tasks like rendering job lists or highlighting specific entries.

This modular approach enhances the user experience by making interactions more precise and understandable for the AI, completing the overview of best practices for MCP app development.