Anthropic MCP has just released the roadmap for the first half of 2025, available at: https://modelcontextprotocol.io/development/roadmap#roadmap.
We have been engaged in research on agent communication protocols and are very focused on MCP, having also proposed some of our suggestions within the community.
Today, combining our understanding of MCP, we will interpret the milestones set by the official team, along with some views on MCP and agent communication protocols.
Milestone Interpretation
Remote MCP Support
I previously mentioned in an article (Why the Recently Released MCP Protocol by Anthropic Should Use DID for Identity Authentication) that the lack of support for remote data access in MCP is due to a fatal flaw — there is no complete identity authentication scheme yet.
It can be seen that the community also recognizes this point, and the first milestone is to support remote MCP connections, with the first requirement being authentication and authorization.
Currently, someone has submitted a proposal for identity authentication, which is a scheme using Oauth2.0 for authentication, and it is under review. We have also communicated with the maintainers in the community, and it is indeed the most realistic solution at this point, as the standardization and widespread application of Oauth2.0 is quite good.
The best part of this proposal is that it uses Oauth2.0 as a standard process while proposing aplugin-based identity authentication scheme. This allows other authentication schemes to be included as experimental proposals in the specification. This greatly helps the flexibility and openness of the specification.
Based on this proposal, we have developed aW3C DID-based identity authentication scheme (https://github.com/chgaowei/AgentNetworkProtocol/blob/main/03-did%3Awba%20Method%20Design%20Specification.md), and we are waiting to submit a PR after the plugin-based scheme is integrated. For the differences between the W3C DID identity authentication scheme and Oauth2.0, you can refer to this article:The Most Suitable Identity Authentication Technologies for Agents: A Comparison of OpenID Connect, API keys, did:wba
The other two aspects of remote MCP support are:
-
Service Discovery, which implicitly includes service descriptions. They may create a product similar to OpenAI GPTs, where the MCP client can call an interface to find its preferred server. -
Stateless Operations, mainly to reduce the development cost of servers.
Distribution and Discovery
This is the third milestone, but I believe it is more important than the second one.
The biggest problem with MCP currently isvery poor usability. Compared to OpenAI’s GPTs, MCP’s capabilities are quite strong, providing developers with powerful means to interact with LLMs. However, the usability is significantly worse than that of GPTs. Currently, users are generally required to have a certain level of development or computer skills, needing to install a bunch of software and services themselves. This greatly raises the threshold for using MCP.
The four items of this milestone, including management, installation tools, sandboxing, and server registration, can indeed make it easier for users to use MCP.
In my vision, for MCP to become widespread, it must lower the usage threshold; while lowering the usage threshold, it also needs to interact with the community ecosystem, requiring a mechanism similar to browser plugins that allows users to easily load different MCP servers. These four tasks should also be aimed at this goal.
Additionally, to improve usability, it is essentialto increase the proportion of remote MCP servers. Anything that can be done remotely should not be done locally. This is also why remote servers are important.
Agent Support
There are three items here: hierarchical agent systems, interactive workflows, and streaming results. The community has discussed this area less; apart from the last item, I do not yet understand the other two well. I will continue to pay attention; if you have a good understanding of this area, feel free to leave a message.
A More Official Ecosystem
MCP hopes that the community can lead the development of standards, although the main maintainers are currently developers from Anthropic. They want to foster a collaborative ecosystem where all AI providers can help shape MCP into an open standard through equal participation and shared governance, ensuring it meets the needs of various AI applications and use cases. They also hope to work with standardization organizations to promote standardization.
This is a good idea and a common practice for building industry standards. Currently, the MCP community is still primarily led by Anthropic, including the review of proposals. Introducing more partners and diluting the dominance of a single vendor can help the specification be adopted by a broader audience. We will also participate more deeply in this in the future.
What Kind of Protocol Do We Need in the Future
Whether agents need a standard protocol for communication in the future seems no longer up for debate. People have gradually recognized its necessity; a recent article from Tencent Technology’s AI Future Guide (Can Mobile AI Help You Order Coffee, Do Apps Agree?) also mentioned this point. Through our communications with industry insiders, we have found that this has gradually become a consensus.
What may still be uncertain is,what kind of protocol we will need in the future? And how will it be accepted by the industry?
We have some thoughts on the outline of future protocols (What Are the Differences with the Agentic Web). We are also designing our ideal agent communication protocol, the AgentNetworkProtocol (ANP): https://github.com/chgaowei/AgentNetworkProtocol.
We believe that future agents must beinterconnected, communicating with the internet through underlying data (API or Protocol), and existing as equals among agents. This is the core principle behind our design of ANP.
Returning to MCP, the biggest difference between the ANP we designed and MCP is in the difference in worldview:
-
MCP is model-centric, with the entire internet being its context and tools -
We (Agent Network Protocol) are agent-centric, where each agent has equal status, forming a decentralized agent collaboration network.
We are on a path that is not particularly the same as MCP; regardless of whether it can be successful, we hope to explore a possibility for the industry.
Finally, if you are also interested in our protocol, feel free to contact us. In particular, if you are a product or developer of AI mobile phones, AI assistants, or agent frameworks, I would love to discuss our views on the future with you.
Feel free to add me on WeChat: flow10240