From: aidotengineer
The roadmap for mCP involves several exciting developments aimed at enhancing its capabilities, promoting wider adoption, and simplifying the development and discovery of AI applications and agents [01:11:14].
Remote Servers and Authentication
A significant upcoming feature is the support for remotely hosted mCP servers, addressing a primary request since the protocol’s launch [01:15:05] [01:15:07]. This means servers can live on a public URL, removing the need for users to understand or host them locally, making them accessible “like a website exists” [01:15:12] [01:15:40] [01:15:49].
The protocol now supports OAuth 2.0 for authentication [01:13:50] [01:13:59]. The server orchestrates the OAuth handshake, guiding the user through the authentication flow (e.g., opening a browser for consent) and then securely holds the OAuth token [01:14:22] [01:14:31]. This allows the server to federate interactions with the end application by providing the client with a session token for future use [01:14:55]. This design prioritizes the server’s control over the interaction with the final application, especially given that clients and servers may not have prior knowledge of each other [01:17:30] [01:17:34].
Registry and Discovery
A major focus is addressing the current fragmentation in mCP server discovery [01:22:00] [01:22:21]. An official mCP registry API is under development, designed as a unified and hosted metadata service [01:22:30] [01:22:38]. This registry will operate in the open, with its schema and development fully public [01:22:41].
The registry aims to solve several problems:
- Discovery: Centralizing how users find mCP servers [01:23:50].
- Protocol Details: Providing information on server protocols (e.g., standard I/O vs. SSE) [01:23:15].
- Trust and Verification: Identifying who built a server and if it’s officially verified by a company (e.g., Shopify’s official mCP server) [01:23:25] [01:23:27]. Users should be judicious about connecting to servers, and the registry will help with trust [01:18:06] [01:18:28].
- Versioning: Supporting versioning for servers, logging changes like new tools or updated descriptions [01:24:04] [01:24:07] [01:20:32].
- Self-Evolving Agents: Enabling agents to dynamically discover and use new capabilities and data on the fly, without needing to be pre-programmed with all possible tools [01:36:01] [01:37:08]. An agent could search the registry for a specific server (e.g., Grafana) and then install/invoke it to perform tasks [01:36:34].
Companies will also be able to host their own private registries, similar to Artifactory [01:24:45] [01:24:47].
Well-Known mCP JSON
Complementing the registry, the concept of a well-known mCP.json file is being explored [01:39:24] [01:40:07]. This standardized file (e.g., shopify.com/.well-known/mCP.json) would provide a verified interface for an agent to discover mCP endpoints, their capabilities, and authentication methods (like OAuth 2.0) for a given domain [01:39:27] [01:39:48]. This allows for a “top-down” approach to discovery, where an agent can directly look up a domain for its mCP services [01:40:17].
This feature is particularly powerful when combined with “computer use” models (like Anthropic’s October 2023 release) that can interact with UIs without APIs [01:40:35] [01:40:39]. An agent could use well-known mCP.json for predefined API interactions and then fall back to computer use (clicking buttons, logging in) for long-tail scenarios without dedicated APIs [01:40:51] [01:41:03].
Medium-Term Considerations
Beyond the immediate roadmap, several areas are being considered for future development:
- Stateful vs. Stateless Connections: Exploring a bifurcation where basic client-initiated requests use short-lived connections (like HTTP), while advanced capabilities (e.g., sampling, server-to-client notifications) use long-lived connections (like SSE) [01:41:41] [01:42:06] [01:42:26]. This would allow clients to disconnect and reconnect without losing context [01:41:53].
- Streaming: Developing first-class support in the protocol for streaming multiple chunks of data from the server to the client over time [01:42:41].
- Namespacing: Addressing tool name conflicts when multiple servers are installed by providing native protocol support for logical grouping of tools [01:42:54] [01:43:00]. The registry will also help with this [01:43:11].
- Proactive Server Behavior/Elicitation: Improving patterns for servers to initiate actions, such as asking the user for more information or notifying them about changes, possibly driven by events or deterministic systems [01:43:31] [01:43:38]. This includes server-initiated notifications when resources change [01:30:06].
Vision for the Future
The combination of sampling (server requesting LLM inference from client) and composability (applications acting as both mCP clients and servers) is particularly exciting for the future of agents [01:11:43] [01:12:00]. This enables the creation of hierarchical systems of agents, where different layers specialize in tasks and can federate sampling requests through the system [01:12:32] [01:12:35]. This “connectivity layer” allows agents to interact with publicly hosted or custom-built servers while maintaining desired privacy, security, and control [01:12:40] [01:12:46]. The vision is for mCP to be the foundational protocol for agents broadly [02:26:36].