Agentic IDEs and Model Context Protocol applied to Mercado Pago
Agentic integrated development environments (agentic IDEs) are reshaping software development by embedding AI-powered assistance directly into coding workflows.
These advanced tools accelerate development cycles, automate debugging, simplify deployment, and improve the overall coding experience. Agentic IDEs enable coordinated modifications across multiple files and significantly reduce manual effort for developers.
At Mercado Libre, over 10,000 developers now use these intelligent IDEs, leading to a significant increase in AI-generated code.
The rise of these tools has also introduced a new conversational coding style called “vibe coding” — a term coined by Andrej Karpathy (see tweet). In this collaborative approach, developers guide and provide feedback while the IDE generates code.
The Model Context Protocol (MCP) enhances this interaction by enabling smooth integration between language models, external tools, and relevant contextual data. This functionality lets developers access a wide range of resources and capabilities within their AI applications. For example, our internal developers use MCPs built for Fury (our IDP) to streamline development providing relevant resources and integrated tooling at their fingertips. Ultimately, MCP creates a more versatile and interconnected development landscape.
As the development world increasingly adopts AI-driven solutions, we’ve been exploring how agentic IDEs, “vibe coding,” and MCP can transform the Mercado Pago integration experience.
Specifically, we aim to address these key questions:
- Can these technologies simplify integration processes and reduce development time?
- How can they empower developers to build more seamless and efficient payment solutions?
- What specific challenges does this approach solve?
The challenge
Integrating payment solutions usually demands significant time and specialized knowledge. Developers must review extensive documentation to understand integration options and technical details. After that, they build the integration using the provided SDKs and APIs. While Mercado Pago continues to improve this process, some of these inherent complexities may still be present.
For this experiment, we focused on one specific integration method: Mercado Pago Checkout Pro. Our goal was to use agentic IDE tools to speed up the integration, aiming for a cost-effective solution in under 30 minutes.
Our approach
Our strategy to address this challenge focused on the following key points, which we’ll explain in detail in the sections below:
- Building a test environment to run our experiments — a demo web app.
- Crafting optimized prompts with specific instructions to start conversations with the IDE.
- Providing access to Mercado Pago documentation and context through MCP.
- Running experiments on integration creation and evaluating results across several agentic IDEs, including Cursor, Windsurf, and Cline.
Test environment
As our test environment, we built a simple web store application called demo-pet-store. This app included:
- Mock product listings
- A functional shopping cart
- A standard checkout flow that required payment integration
This web store served as our baseline for building the integration with Mercado Pago across multiple agentic IDEs.
Optimized prompts
To make sure the agentic IDEs have the right context for the integration task, we developed an enhanced prompt with detailed instructions. We optimized this prompt using Claude 3.5-sonnet:
Documentation and context through MCP
The Model Context Protocol (MCP) defines a client-server architecture, where:
- Hosts are LLM (large language model) applications, such as agentic IDEs, that initiate connections.
- Clients maintain one-to-one connections with servers within the host application.
- Servers are external programs that provide resources (such as file contents, database records, or API responses), tools (functions or actions callable by the agents), and prompts (reusable prompt templates for specific tasks) to clients.
To provide enhanced documentation and context during the integration process, we built an experimental MCP Server for Mercado Pago: demo-mercadopago-mcp-server.
This server includes a tool called search_documentation that lets you search the Mercado Pago Developers Site documentation. Behind the scenes, this tool integrates with the Mercado Pago Developer Site’s Search Docs API to provide helpful context for different parts of the Mercado Pago integration.
The image below shows an example request/response for this tool. For further details about the parameters, see the README:
Experiments and results on integration creation
We tested our approach across three agentic IDEs: Cursor, Windsurf, and Cline.
For each agentic IDE, we:
- Loaded the demo-pet-store project as our initial test environment.
- Configured the Mercado Pago MCP Server (demo-mercadopago-mcp-server) by following its installation instructions.
- Started a new agent conversation using the optimized prompt.
- Provided feedback until the task was complete.
In all cases, we achieved fully functional integrations of Mercado Pago Checkout Pro, including redirecting to the MP Checkout, handling the return-to-back URL flow, and setting up webhooks, as shown in the videos below:
Cursor + Mercado Pago MCP Server
Windsurf + Mercado Pago MCP Server
Cline + Mercado Pago MCP Server
Key findings
Our experiments with AI-assisted Mercado Pago integration revealed several interesting insights:
- Documentation quality is crucial: While the search_documentation tool, integrated with the developer site Search API, offers a general, scalable, and maintainable solution, improving the quality of the documentation is essential. High-quality documentation provides clearer and more reliable context for agentic IDEs.
- AI behavior patterns: We observed that AI agents often make multiple queries to find the expected information, particularly code examples in specific technologies. When responses are not satisfactory — as in the case of Windsurf — the IDE begins to explore less reliable alternatives, such as searching the web for additional information.
- Functional results: Despite some challenges, all tested environments successfully integrated Mercado Pago Checkout Pro, achieving impressive implementation times of less than 30 minutes.
Conclusion and next steps
Our experiments show that AI-assisted development can significantly accelerate Mercado Pago integration, with fully functional implementations completed in under 30 minutes in all the agentic IDEs tested.
Using the Model Context Protocol (MCP) to provide context and tools proved to be an effective approach and worked across all the agentic IDEs without any needed changes.
While there’s still room to improve documentation and search capabilities, this approach is promising for developers looking for a “vibe coding” experience when implementing payment solutions in their applications. However, developers should still understand the underlying code to reduce risks from over-reliance on these advanced tools and to keep their applications robust.
The Mercado Pago Official MCP Server is now live, enhancing developer interactions and integrations. Learn more in the official documentation.
