Beyond request-response: How MCP servers are learning to collaborate

The Model Context Protocol (MCP) initially followed a straightforward request-response model where servers expose tools and large language models (LLMs) decide when to use them, but as it entered production, limitations emerged, prompting extensions to enhance collaboration while maintaining explicit control. The protocol introduced collaborative patterns like sampling, URL-mode elicitation, and form-mode elicitation to address issues such as handling sensitive authentication flows, resolving ambiguities, and incorporating server intelligence without compromising user input and oversight. Sampling allows servers to engage LLMs for reasoning and validation, URL-mode elicitation ensures secure interactions outside the protocol for sensitive tasks, and form-mode elicitation clarifies ambiguities with structured user input. An ongoing debate within the MCP community explores the potential for bidirectional tool calls, which would allow servers to proactively initiate interactions, raising questions about trust, consent, and predictability. As MCP evolves, it seeks to enhance structured collaboration without granting full autonomy to servers, gradually shifting towards a model where servers, models, and users collaboratively shape workflows while carefully balancing the risks and benefits of expanded server capabilities.