Table of Contents
As the technology of large language models (LLMs) rapidly advances, the ability of AI to handle long contexts has emerged as a new challenge. For AI to possess true intelligence, it must understand continuous and meaningful context, not just simple data transfer.
In this regard, Anthropic's MCP (Model Context Protocol) is noted as a pivotal solution to the long context processing problem. However, even prior to the emergence of MCP, various tech companies have been making efforts to process long contexts and integrate tools in different ways, and this issue continues to be addressed through various methods.
Thus, the future advancement of AI seems to depend on how effectively it can enhance these long context processing capabilities.
MCP and AI Connection Technology Series
OpenAI's Plugin and Function Calling Features
In 2023, OpenAI unveiled the ChatGPT plugin feature, enabling integration with external APIs. This allows the model to interact with various external tools, such as web browsing and third-party service calls, making it possible to perform specific tasks according to the OpenAPI specification.
However, the plugin method had the drawback of functioning only in the restricted environment of ChatGPT's web interface. Each plugin required individual integration and hosting, which led to limitations in scalability and consistency. Notably, the need for separate authentication procedures for each plugin made maintaining a continuous dialogue flow and tool recognition challenging.
Subsequently, OpenAI added the Function Calling feature, enabling the model to call developer-defined functions in JSON format. This feature, when combined with frameworks like LangChain, supported more structured tool usage. However, this approach also has the limitation of being static, allowing access only to predefined tools.
These changes contribute to enhancing the utilization of AI models, but there are still areas needing improvement. This is why the future development is anticipated.
Limitations Addressed by MCP
MCP is an innovative system that overcomes existing limitations. Thanks to its structure that allows models to automatically recognize and utilize new MCP servers during runtime, dynamic access to tools can occur without predefined definitions.
For example, LangChain has been improved through recent updates to treat MCP servers as a single tool source. This enables the combination of MCP clients and servers with function calling, allowing AI to effectively utilize tools and their results. Such changes enhance the potential for applications across various fields.
Adoption of MCP by Microsoft and OpenAI
MCP is led by Anthropic, with OpenAI and Microsoft also participating in this new standard. OpenAI plans to add MCP support to the agent SDK by the end of 2024, enabling OpenAI model users to easily utilize various tools in the MCP ecosystem.
Meanwhile, Microsoft integrated MCP into its Azure AI Agents service. In March 2025, the Azure blog demonstrated use cases where an AI like Claude could access Bing search or corporate SharePoint documents in real-time using MCP. This indicates that MCP is establishing itself as an open standard compatible with multiple platforms and models. Such developments are expected to broaden the applications of various AI technologies.
Google Bard's Extensions Feature
In 2023, Google added a new Extensions feature to its chatbot Bard. This feature enables the AI to access platforms like Gmail, Google Drive, and Google Docs. As a result, the AI now has the ability to read and summarize documents in a user's mailbox or drive, which shares similar expansion goals with MCP.
However, Google's Extensions operate as a closed integration within the Google ecosystem, which is a stark contrast to MCP’s goal of an open connection that anyone can freely use. This differentiation is expected to play a crucial role in shaping how AI is utilized and the accessibility for users going forward.
Enhancement of Long Document Processing Capabilities of the Model
There is another method to extend the model's context window. OpenAI supports up to 128,000 tokens in GPT-4 Turbo, and the later released o3 model reportedly supports up to 200,000 tokens. Google's Gemini model is said to have the capability to handle over a million tokens.
These advancements represent the possibility of AI understanding a volume of text equivalent to an entire book at once. However, there are clear limitations to simply aggregating all data for processing.
According to research presented at ICLR 2024, just adding search-based context can yield performances similar to large window models. Furthermore, it was confirmed that information retrieval always contributes to enhancing performance, regardless of the size of the context window. These research results are expected to further broaden the potential applications of AI.
The Best Strategy for Handling Long Contexts is 'Connection'
Based on the discussion above, there are several approaches to solve the long context problem.
First, there are closed plugins (OpenAI, Google) that are used within platforms. Second, there is the method of static function calling and code-based tool connections (Function Calling, LangChain). The third approach is through the model's own window extension (GPT-4 Turbo, Gemini). Finally, there is MCP, an open standard that allows for dynamic connections.
Among these, MCP stands out with its unique structure that combines openness and versatility, establishing itself as a de facto industrial standard that can be used across various models and platforms. This structure is expected to further accelerate technological advancement.
Ultimately, while there are various ways to offer AI more context, the core remains the same.
Connecting necessary information for the AI to use it in a timely, safe, and unrestricted manner.
And currently, the technology closest to that answer is
MCP
.
#MCP, #AItoolintegration, #OpenAIplugin, #functioncalling, #LangChain, #AzureAI, #GoogleBard, #GPT4Turbo, #Gemini, #contextwindow, #longdocumentAI, #AIcontextexpansion, #AIecosystem, #AIagents, #AIstandards, #MCPadoption, #Anthropic, #OpenAIagentSDK, #MicrosoftAzure, #cloudAI, #openAI, #AIopenstandard, #AIautomaticintegration, #agentframework, #LangChainMCP, #AIcomplementstrategies, #searchbasedAI, #ICLR2024, #OpenReview, #AIfuturetechnology
