December 2025 Summaries

Emerging best practices for managing Model Context Protocol (MCP) and AI agents in production environments are focusing on controlling complexity, improving security, and centralizing control through enterprise-grade architecture. As MCP transitions from experimental to production use, challenges arise from tool complexity and governance issues, prompting organizations like Microsoft, Anthropic, and Gartner to emphasize efficient context window management, deterministic tool orchestration, and active tool management. Lunar's MCPX platform operationalizes these practices by enforcing context efficiency, introducing composite tools for deterministic orchestration, and managing tool exposure through curated groups. Additionally, Lunar supports tool customization for safe reuse, encourages remote-first MCP consumption with guarded local exceptions, and advocates for gateway-centric governance to centralize authentication, authorization, and observability. As the MCP ecosystem matures, architectural discipline becomes essential, shifting responsibility from agents to infrastructure to ensure reliable and secure MCP deployments.

HiBob, a global HR technology company, successfully scaled its AI adoption by partnering with Lunar and implementing MCPX as a centralized gateway for managing AI agents and MCP usage. As AI tools became integral across various departments, HiBob faced challenges such as security bottlenecks and the absence of a central MCP registry. By integrating MCPX with the company's identity provider, HiBob enabled identity-aware governance, allowing for automatic access control and the creation of department-specific MCP profiles. This approach improved security team efficiency by reducing manual reviews and approval delays, while also providing engineers with faster onboarding and clear guidance on safe MCP usage. The implementation resulted in enhanced visibility, consistent governance, and maintained business velocity, demonstrating that HiBob could safely expand AI utilization without compromising security.

Model-Context-Protocol (MCP) servers, which enable AI agents to access real-world systems, present unique security challenges such as tool poisoning, prompt injection, and context manipulation. As organizations adopt these servers, understanding the threat landscape becomes essential, particularly as attackers can exploit these systems to execute malicious actions, exfiltrate data, or corrupt systems. Lunar's platform, MCPX, incorporates OWASP guidelines to mitigate risks through automated risk-scoring and governance workflows, which include version drift detection, tool description analysis, and the classification of tool sensitivity. By leveraging a large language model, MCPX reviews tool definitions, assesses security risks, and recommends constraints to minimize potential damage. This system classifies tools into risk-based tiers, offering a structured approach to manage and monitor tool usage, thus enabling organizations to safely harness the capabilities of AI agents in production environments while maintaining security and compliance.

AI agents using the Model Context Protocol (MCP) can face significant challenges when connected to an excessive number of tools, leading to prompt bloat, slower response times, increased costs, and a higher risk of errors or unsafe tool usage. This tool overload occurs because large language models (LLMs) struggle with too many unnecessary options, which can confuse reasoning and lead to poor tool selection or even hallucinations of nonexistent tools. The issue is exacerbated by the modular nature of MCP, where tools from various systems can be easily aggregated, often resulting in an unwieldy and inefficient setup. To address this, the concept of "Tool Groups" is introduced, allowing agents to access only the tools necessary for specific workflows, thereby reducing context window consumption and improving overall agent performance. This approach not only enhances agent accuracy and reduces costs but also mitigates security risks and helps manage platform-imposed tool limits effectively, ensuring more efficient and reliable AI operations.