Sara Du, Founder, Asari: MCP & the Future of AI Agents

Speaker Background & Context

• Presenter has nearly six years of experience building B2B SaaS, specifically via "Alloy," an API integration platform that raised $27 million from investors including Andreessen Horwitz, Bain, and YC.
• Formerly a Harvard drop-out and Thiel Fellow, the speaker pivoted to focus on non-deterministic software agents after realizing existing deterministic software interfaces are too rigid for AI tasks.

MCP Definition and Traction

• MCP (Model Context Protocol) was created by Anthropic in November of the previous year.
• Major technology firms including OpenAI, Microsoft, and Ollama have adopted MCP, driving rapid ecosystem growth.
• GitHub repository traction for MCP reached 60,000 stars within one year, outpacing the 10-year adoption curve of OpenAPI, which currently holds approximately 30,000 stars.
• The protocol standardizes how Large Language Models (LLMs) connect to external tools, replacing rigid, pre-programmed API integrations with dynamic context discovery.

Technical Mechanics: API vs. MCP

• Unlike traditional Application Programming Interfaces (APIs) which require deterministic code to interact with specific endpoints, MCP allows models to interpret English descriptions of tools to determine logic on the fly.
• MCP decouples the agent from the need for pre-trained knowledge of specific APIs, enabling LLMs to utilize unfamiliar applications (e.g., NetSuite, Salesforce) by analyzing tool descriptions rather than hard-coded schemas.
• The speaker defines an AI agent as a program that operates autonomously and non-deterministically, distinguishing it from traditional scripted automation.

Adoption and Use Cases

• Current implementation relies on "clients" (e.g., Claude Desktop, Cursor, Windsurf) to interface with "MCP servers," though official support for MCP remains in early stages.
• High-value non-deterministic use cases identified include front-end testing (where human QA variance is beneficial), text generation, and cross-application workflow automation (e.g., moving data from HubSpot to Notion).
• Business applications for MCP include automated CRM cleanup, identifying duplicate workflows, and removing unused custom fields without requiring human RevOps intervention or custom scripts.

Security Risks and Challenges

• A critical security vulnerability was recently exposed when the "MCP Inspector" testing tool was hacked, highlighting the risks of the current "Wild West" open-source environment.
• Malicious MCP servers can embed prompts to steal user credentials when downloaded directly from public repositories like GitHub.
• Enterprises are currently grappling with the need for internal security scans, whitelisting mechanisms, and the creation of private MCP registries.
• Quality variance is high because many open-source servers are simply direct translations of human-centric APIs rather than interfaces optimized for agent interaction.

Strategic Outlook and Future Trends

• The speaker predicts the emergence of new business models where companies operate primarily as MCP servers, similar to how Stripe and Twilio leveraged APIs, without necessarily building traditional web applications.
• Future development will focus heavily on "multi-agent systems," requiring new infrastructure to manage agents as a workforce component.
• Evaluation frameworks are currently underdeveloped; successful deployment requires defining "golden paths" to test agent performance, as agents may take suboptimal numbers of steps to reach a solution.
• Underrepresented adoption sectors are expected to include marketing and operations, moving beyond the current developer-centric focus.