Lecture Notes on mCP and Serverless Functions

Introduction

• Recent exploration of mCP (Modular Component Protocol) and its implementation.
• Initial success with running mCP on a local computer using standard IO transport.
• Desire to implement mCP as a REST API for easier deployment and invocation.

Discovery of mCP Limitations

• Stateful Protocol: mCP requires a stateful server connection, unsuitable for simple REST APIs.
  • Needs Server-Sent Events (SSSE) and a long-lived connection.
• Serverless Environment Challenges: Incompatible with serverless environments, which are ideal for quick, scalable functions.
• Deployment Complexity: Stateless REST APIs are common; mCP complicates deployment by requiring stateful configurations.

mCP Features and Issues

• Sampling Feature: Allows server requests for AI tasks, but introduces security concerns.
  • Potential risk with AI agents having access to API keys.
  • No existing mCP clients support this feature due to lack of incentive.
• State Requirement Justification: Mainly for sampling and notifying clients of changes.
  • Alternatives: Client polling and error handling for changed resources.

Criticism of mCP

• Over-engineered: Protocol deemed too complex for intended use.
• Suggested Simplification: One-way communication preferred, with AI agents making requests and servers responding.

Alternative Solutions

• Wildcard AI's agents.json: Proposes a simpler method using JSON atop Open API specs.
  • Provides endpoint descriptions without changing server deployments.
• Adoption Potential: Lightweight, leverages existing technologies, and standardizes AI-agent interactions.

Community Feedback and Observations

• Context Window Overload: Registering many tools with AI agents can overwhelm and confuse the agent.
• Routing Component Suggestion: Enables dynamic tool listing based on relevance to task.

Personal Outlook

• Future Work: Unlikely to pursue mCP further unless a stateless version is developed.
• Interest in Alternatives: Open to exploring alternatives like agents.json for better integration with serverless functions.

Conclusion

• Call to action for community suggestions on alternatives.
• Emphasis on finding transactional exchange solutions that fit serverless models.