Theodora “Theo” Chu is a product builder who helped shape the Model Context Protocol, an open standard connecting AI systems to external information, tools, and services. Her work focuses on a central challenge for useful AI agents: giving models access to the context and capabilities they need to act effectively beyond a chat window.
At Stanford, Chu co-authored research on detecting abusive online comments, comparing deep-learning approaches and finding particular promise in character-level representations of messy language. She subsequently built products at Stripe, founded an AI startup, and worked on Anthropic’s API and MCP.
MCP emerged from engineers David Soria Parra and Justin Spahr-Summers’s frustration with manually transferring information from applications such as Slack and Sentry into language models. Following its open-source release in November 2024, adoption by Cursor, Visual Studio Code, and Sourcegraph helped turn the protocol into a shared interface for AI integrations. Chu’s contribution centers on the product decisions and ecosystem infrastructure needed to make that interoperability practical.
- Three users of every AI tool. Integrations must serve the person requesting help, the developer building the client, and the model choosing which action to take. Chu argues that exposing existing API endpoints verbatim often produces poor tools; effective interfaces reflect user intentions and the context models require.
- Server-first protocol design. Expecting servers to outnumber clients, Chu favors making servers easier to build even when clients assume additional complexity. This approach extends to authentication, developer SDKs, debugging, Streamable HTTP, and elicitation, which allows servers to request clarification from users.
- Security for agents with real-world access. Chu sees opportunities in better integrations, hosting, evaluation, observability, and auditing, especially beyond developer tools in fields such as finance, law, and education. Her account of MCP’s development also anticipates automated server generation, while emphasizing that greater access to real data and external actions increases privacy and security risks.
