MCP Is Not Good Yet — David Cramer, Sentry

Key Concepts

• MCP (Managed Code Program): A pluggable architecture for agents.
• OAUTH 2.1: An authentication protocol.
• Agents: Services with a new name, leveraging LLMs for specific tasks.
• Context: Providing relevant information to agents for effective operation.
• Remote MCP Server: Focus on OAUTH specification for integration into various agents.
• Standard IO: An interface that is not super useful for businesses.

MCP as a Pluggable Architecture

• MCP is defined as a pluggable architecture for agents, suitable for enterprise cloud services.
• The core idea is to enable collaborative bug fixing, where tools like Sentry and Cursor can work together.
• The speaker emphasizes that MCP is not just about exposing existing APIs but requires designing around the system to provide relevant context to agents.

Implementation Challenges and Solutions

• OAUTH Complexity: Implementing OAUTH 2.1 can be challenging due to limited support. Century uses Cloudflare Shim to proxy OAUTH 2 API on top of Cloudflare workers.
• JSON Payload Reasoning: Robots (agents) struggle with giant JSON payloads not designed for them.
• Client Compatibility: Reliance on clients (VS Code, Cursor) for native authentication and stability.
• Error Handling: Designing human-readable error messages is crucial because the agent reasoning about the API is abstract.

Remote vs. Standard IO Interface

• The speaker advocates for focusing on the remote MCP server and OAUTH specification, especially for B2B SaaS companies.
• The standard IO interface is deemed less useful and poses significant security risks, particularly prompt injection.
• Security concerns are highlighted, emphasizing the need to trust only reputable MCP tools and avoid downloading random packages.

Practical Learnings and Recommendations

• OAUTH Focus: Only care about OAUTH if you're a B2B SaaS company.
• Context is Key: You have to spend the calories and can't just proxy Open API and expose it as tools.
• Markdown for Responses: Returning markdown is suggested as a way to provide structured information that both humans and language models can easily understand.
• Agent Design: Focus on building agents and optimizing for context in workflows.
• Cost Considerations: Be mindful of the cost implications of passing large amounts of data, as it can significantly increase API call expenses.

Examples and Use Cases

• VS Code Integration: Century's MCP integration with VS Code allows users to look up data from Sentry and potentially fix bugs more efficiently.
• Bug Fixing Example: A real-world example of using VS Code with Century's MCP to automatically fix bugs, highlighting the potential and current limitations.
• Century's Seir Agent: An example of an agent that performs root cause analysis of application errors, exposed through both the UI and MCP.

Agent-Centric Approach

• The speaker argues that the real value lies in building agents, which are essentially services with a new name.
• Agents provide more control over prompts, results, and the overall workflow, enabling better optimization and responsibility.
• The future of B2B is seen in exposing agents through the MCP architecture, treating MCP as a plug-in architecture.

Technical Details

• OAUTH 2.1: A specific version of the OAUTH protocol that presents implementation challenges.
• Cloudflare Shim: A technology used to proxy OAUTH 2 API on top of Cloudflare workers.
• Token Limits: Constraints on the amount of data that can be passed in API calls, affecting the design of MCP tools.
• Streaming Responses: The lack of streaming responses for tools is identified as a significant limitation for agent-to-agent communication.

Conclusion

The speaker advocates for a strategic approach to MCP, emphasizing the importance of building agents, providing relevant context, and carefully considering security and cost implications. While acknowledging the current limitations and challenges, the speaker expresses optimism about the future potential of MCP as a pluggable architecture for enabling collaborative and intelligent workflows. The key takeaway is that MCP is not a simple add-on but requires a fundamental shift in thinking towards agent-centric design and context optimization.