Theodora Chu

Theodora Chu worked on Model Context Protocol (MCP) as a product manager at Anthropic in 2025. Her public work included explaining the protocol’s design decisions, advising founders on building useful MCP servers, and coauthoring the September 8 MCP Registry preview announcement. In her AI Engineer World’s Fair 2025 talk, she introduced herself as Theo.[3][2]

Chu argued that an MCP server has three users: the end user, the client developer, and the model. She criticized exposing existing API endpoints one for one as tools. Her alternative was to work backward from the end user’s use cases and likely prompts, then decide which capabilities the model needs to respond correctly. This makes the model’s ability to use an interface part of the design problem.[3]

She explained a deliberate team trade-off: expecting many more servers than clients, the team prioritized server simplicity and better server-building tools, accepting additional complexity for clients. Chu explicitly acknowledged that the prediction could be wrong. She also described the move to Streamable HTTP as controversial, linking the decision to the team’s expectation that agents would need to communicate with one another.[3]

Chu publicly acknowledged that the team had initially got authorization wrong: “We fixed Auth”. At the time of the talk, the correction was in the draft specification. She credited community contributors who explained how authorization needed to work with identity providers and helped update the protocol.[3]

Her example of elicitation, then another draft feature, showed why an agent sometimes needs clarification rather than another action. A flight-booking server receiving a request for the best flight to Atlanta could ask whether “best” meant cheapest or fastest, then receive the user’s answer through the client. The example illustrated gathering a missing decision criterion, without specifying a chosen flight or completed booking.[3]

In her World’s Fair talk, Chu discussed registry API development as a way to help models find MCP servers they had not been given upfront. On September 8, 2025, she coauthored Introducing the MCP Registry with David Soria Parra, Adam Jones, Tadas Antanavicius, and Toby Padilla. The launch introduced an open catalog and API for publicly available MCP servers.[3][2]

The registry supplied maintainer-reported server information for downstream public marketplaces and private enterprise catalogs to adapt. The coauthors emphasized preserving existing community and company registries rather than replacing them. They credited Antanavicius with spearheading the initial effort alongside Alex Hancock, followed by maintainers Padilla and Jones and wider community contributions. At launch, the preview offered no data durability guarantees and could undergo breaking changes.[2]

Chu was also acknowledged, alongside John Welsh, David Soria Parra, and Adam Jones from MCP, as a contributor to Ken Aizawa’s September 11, 2025 article Writing effective tools for agents — with agents. The article recommended workflow-oriented tools, realistic agent evaluations, and iterative optimization checked against held-out tests. Her individual contribution was not specified.[4]

Chu’s 2025 advice to founders emphasized more and higher-quality MCP servers, including applications beyond developer tools in sales, finance, legal work, and education. She assigned an illustrative 80/10/10 weighting to better servers, simpler server-building tooling, and security tooling—her priorities for builders, rather than measured market shares. Hosting, testing, evaluations, and deployment were opportunities for both internal enterprise servers and products serving external users.[3]

She framed automated MCP server generation as a speculative “moonshot”: increasingly capable models might eventually write integrations on demand. Her more immediate call concerned security, observability, and auditing. She argued that external access and real data increase AI applications’ security and privacy implications, and stressed that this need extends beyond MCP.[3]

Before her MCP product role, Chu had been an AI startup founder, according to her introduction at the World’s Fair. That earlier experience provides context for her direct advice to founders building in the protocol’s ecosystem.[3]

In recounting MCP’s beginnings, Chu credited David Soria Parra and Justin Spahr-Summers with the initial idea in mid-2024. They were repeatedly pasting external context—such as Slack messages and Sentry error logs—into a model’s context window. In her account, the resulting ambition went beyond retrieval to letting models interact with external systems. She described internal workflow experiments at Anthropic’s November 2024 Hack Week, followed by open sourcing that month, and argued that adoption depended on builders finding the protocol useful in their own work.[3][2]