After OpenAI launched the image generation feature of GPT-4o yesterday, which received widespread praise from the community, they announced another significant development today: their open-source agent tool Agents SDK now supports MCP (Model Context Protocol), marking a crucial step in AI model interoperability.
This update immediately sparked heated discussion in the community, demonstrating the industry's focus on MCP and viewing OpenAI's support as a major endorsement of this standard.
What is MCP? Giving AI an "External Brain" and "Operational Hand"
MCP, short for "Model Context Protocol", was initially proposed and open-sourced by Anthropic in November 2024, with the core goal of creating an integrated interface for AI models to access external data and tools.
Traditional language models like ChatGPT, despite their powerful language capabilities, cannot directly access user devices, databases, or third-party services. MCP designs a standardized communication method allowing large language models (LLMs) to safely read external data (like an Excel spreadsheet), use tools (like GitHub's API), and even perform actions independently (like running a Linux command).
Developers describe MCP as the "USB slot for AI". It evolves AI from a mere conversational virtual assistant to an action-capable agent that can execute tasks.
OpenAI Supports MCP
Now, OpenAI announces that its Agents SDK supports MCP, enabling developers to directly integrate external services into the Agents system through the MCP protocol, implementing functions like data reading, API operations, and command execution. This significantly expands the practicality of OpenAI agents, such as integrating web search, professional analysis, and local queries.
OpenAI simultaneously previewed future MCP support expansion, including:
Opening OpenAI API direct MCP support;
Adding MCP functionality to ChatGPT desktop version. This might include a frontend "MCP tool marketplace" where users can select desired tools or services (like a Notion MCP plugin).
This move means not only can developers more flexibly build AI agents, but ordinary users will also directly benefit from the "all-purpose assistant" experience brought by MCP.
Agents SDK is Open Source and Releases MCP Implementation Method
Notably, OpenAI Agents SDK uses the MIT license, the most open free software authorization type. Developers can freely modify, commercialize, and redistribute. The newly added MCP support is also an open-source project, allowing global developers to create intelligent agent services meeting MCP standards and expanding AI application boundaries.
Specifically, the official GitHub repository link introduces two types of MCP servers:
Standard Input/Output (stdio) Server: Runs as a subprocess of an application, similar to local execution.
HTTP-based Server: Connected via URL, suitable for remote servers.
Additionally, it provides examples of integrating MCP servers into Agents, along with information about tool list caching and tracing, offering comprehensive technical support for developers implementing MCP in Agents SDK.
Other LLM Support for MCP
Currently, platforms supporting MCP include Anthropic's Claude model, the open-source multi-model chat platform LibreChat, and the developer-oriented editor Zed.
Now, with OpenAI joining the MCP ranks, it indicates that this standard is becoming increasingly popular and might become the "universal plug" of the AI industry, making collaboration easier between different AIs and tools, and potentially attracting Google Gemini, Meta LLaMA, xAI Grok, and other LLMs to officially support MCP.
