1 : Service Informatique : Développement, Exploitation et Assistance Laboratoire d'Analyse et d'Architecture des systèmes
In this work, we present an Agentic Web Platform (AWP), a general framework for building AI-driven web applications in which large language models (LLMs) operate strictly through a controlled tool-based interface. At the core of the platform is the use of the Model Context Protocol (MCP) as an enforcement layer that constrains LLM behavior to a predefined set of capabilities exposed by MCP servers. This design ensures that all application logic is executed through auditable, secure, and well-defined tool invocations, enabling reliable and governable agentic systems. To demonstrate the platform's capabilities, we introduce an agentic web form application as a representative use case. This application showcases how users can specify requirements in natural language - e.g., “Add a text field called ‘Email' after the name field” - and have the system automatically construct and manage the corresponding web interface. Unlike traditional approaches, the LLM does not directly generate or manipulate application code; instead, it plans and executes tasks exclusively via MCP-exposed tools, ensuring correctness, traceability, and alignment with backend constraints. The platform adopts a multi-agent architecture in which specialized agents collaborate through MCP servers to handle planning, execution, and state management. A key component is the Plan Manager, which implements a “Think-First” protocol to decompose user requests into structured, verifiable task sequences. These tasks are then executed through domain-specific MCP servers that act as intermediaries between the agents and the underlying web services. This separation enforces a strict boundary between reasoning and action, improving robustness and reducing the risk of unintended behavior. Architecturally, AWP follows a layered design that decouples presentation, application logic, and data management. The backend, implemented with FastAPI, exposes a set of APIs for resource manipulation and persistence, while MCP servers encapsulate these APIs as callable tools with explicit schemas and constraints. Security and governance are central to the platform's design. By enforcing that all LLM actions pass through MCP servers, the system achieves fine-grained control over permissible operations. Additional safeguards include authentication, rate limiting, input validation, and comprehensive audit logging. A Human-in-the-Loop (HITL) mechanism is integrated for high-risk or irreversible operations, requiring explicit user approval before execution. Together, these measures ensure that agentic behavior remains transparent, controllable, and aligned with user intent. The technology stack combines modern, production-ready components, including FastAPI and MongoDB for backend services, React for the frontend, and FastMCP for tool orchestration and state management. Deployment is supported through containerized environments and secure communication channels with TLS encryption. In summary, the proposed Agentic Web Platform generalizes the concept of AI-assisted application development by introducing MCP as a foundational control layer for LLM-driven systems. The web form application illustrates how this paradigm enables flexible, user-friendly, and secure application generation while maintaining strict operational guarantees. This approach paves the way for a new class of agentic web applications in which LLMs act as planners and coordinators, rather than unconstrained executors. Future work will explore broader application domains, richer tool ecosystems, and advanced coordination strategies among agents.
