Why We Chose MCP for Our AI Agent and Why It Was a Good Decision

Michael Vasiliev, a developer at Bitrix24, describes why the company chose MCP (Model Context Protocol) for its AI assistant, which started as Marta and evolved into BitrixGPT. Initially, the assistant used function calling, which was sufficient for a few fixed actions but became cumbersome as the number of tools grew, requiring constant code updates and loading all tools for every user. MCP allowed tools to be moved outside the agent into an external service, enabling dynamic tool availability based on user permissions. The first version was internal, using an MCP proxy with signed user IDs for authorization. To support external clients, they added OAuth and switched from SSE to Streamable HTTP transport. MCP was officially introduced by Anthropic in November 2024, and Bitrix24 adopted it in March 2025. The early adoption had advantages, like building architecture around a protocol that later improved, but also challenges, like adapting to protocol changes. They also introduced virtual MCP servers for different domain areas (tasks, calendar, CRM) and a mcp-cli to avoid Tool Hell, which occurs when too many tool descriptions clutter the model's context. Long tool responses are saved to files to prevent context overload, and overall MCP helped build a scalable infrastructure for tool management.