Model Context Protocol: Origins and Requests For Startups — Theodora Chu, MCP PM, Anthropic

Key Concepts

• Model Agency: Giving Large Language Models (LLMs) the ability to interact with the outside world and make decisions.
• MCP (Model Communication Protocol): An open-source, standardized protocol for LLMs to access external tools and data.
• Servers: Components that expose functionalities or data to LLMs via MCP.
• Clients: Applications or platforms that use LLMs and MCP to interact with servers.
• Agents: LLMs that can autonomously choose actions and interact with the environment.
• Streamable HTTP: A transport protocol enabling bidirectional communication between clients and servers, facilitating agent-to-agent communication.
• Elicitation: The ability for servers to request additional information from end-users through the client.
• Registry API: A mechanism for models to discover MCPs that weren't explicitly provided to them.

Origin Story and Motivation

• The co-creators of MCP, David and Justin, observed the common practice of copying and pasting context from external sources (e.g., Slack, Sentry) into LLM context windows.
• The core idea was to enable LLMs to "climb out of their box" and access real-world context and actions.
• The goal was to address the limitation of model agency, which was seen as a major obstacle to the usefulness and intelligence of LLMs.
• The decision to make MCP open-source was driven by the need to avoid closed-source ecosystem limitations, BD/partnership dependencies, and the complexities of aligning interfaces.

Development and Adoption

• A small team developed the initial MCP protocol and launched it at Anthropic's hack week in November of last year.
• The internal launch was successful, with engineers building MCPs to automate their workflows.
• MCP was open-sourced in November of last year.
• Initial reception was mixed, with questions about the need for a new protocol and the value of open-source.
• Adoption by Cursor and other coding tools was a turning point, enabling builders to create MCPs for themselves.
• Later adoption by Google, Microsoft, OpenAI, and others further solidified MCP's position.

Principles and Design Decisions

• The primary goal is to enable model agency and the development of agents.
• Agents are defined as systems where the model's intelligence is used to choose actions.
• The protocol supports streamable HTTP for bidirectional communication, which is crucial for agent-to-agent interactions. This was a controversial decision.
• The design prioritizes server simplicity, even if it increases client complexity, based on the belief that there will be more servers than clients.

Project Updates

• Launched the ability to build remote MCPs.
• Fixed OAuth implementation based on community feedback.
• Adopted streamable HTTP as the primary transport.
• Made updates to SDKs and the Inspector debugging tool to improve developer experience.

Future Directions

• Focus on improving the agent experience.
• Added elicitation to the draft spec, allowing servers to request more information from end-users.
• Developing a Registry API to enable models to discover MCPs dynamically.
• Investing in open-source examples to establish best practices and standards.
• Working on the next phase of governance to ensure MCP remains open.

Building Agents with MCP

• An agent is essentially a server acting as a client and vice versa, enabling communication with other agents, servers, and clients.

Opportunities in the MCP Ecosystem

• Build More Servers: The ecosystem needs more high-quality servers for various verticals beyond dev tools (e.g., sales, finance, legal, education).
  • "A lot of people are wrapping their API endpoints one to one and just exposing that as tools. I don't think that's the right way to build an MCP server."
  • When building a server, consider the needs of the end-user, the client developer, and the model itself.
• Simplify Server Building: Create tooling to make it easier for enterprises and indie hackers to build MCP servers (e.g., hosting, testing, eval, deployment).
• Automated MCP Server Generation: Explore the possibility of models automatically writing MCPs on the fly as their intelligence improves.
• AI Security and Observability: Develop tooling for AI security, observability, and auditing, especially as AI applications gain access to more real-world data.

Conclusion

MCP aims to unlock the potential of LLMs by enabling model agency and facilitating interaction with the external world. While still early, the ecosystem presents significant opportunities for building servers, simplifying server development, and addressing security concerns. The future of MCP hinges on community involvement and the continued evolution of model intelligence.