MCP began as an open-source protocol for giving models access to context and actions outside their context window.
2
MCP server builders need to design for the end user, the client developer, and the model, rather than simply expose API endpoints one for one.
3
The largest startup opportunities are higher-quality servers, server-building tools, automated MCP generation, and security, observability, and auditing products.
Summary
Theodora Chu explains why Anthropic created MCP and how its purpose extends beyond tool calling. The protocol grew from a practical problem: engineers were copying context from Slack, Sentry, and other systems into model prompts. An open standard could let models reach those systems directly and give more builders access to the ecosystem. Chu describes design choices such as streamable HTTP, server simplicity, elicitation, registry work, and community-led changes to authorization. Her main request to startups is to build more high-quality servers for industries beyond developer tools. She also wants better hosting, testing, evaluation, and deployment tools for server creators. Chu expects models may eventually generate MCP servers dynamically. As models gain access to real data and external actions, she sees demand for security, observability, and auditing products. She is candid that the ecosystem is early and that some protocol decisions may change.
MCP started as a way to stop engineers copying context into model prompts
The co-creators of MCP, David and Justin, noticed that their work repeatedly involved moving useful information from outside systems into a model's context window. A Slack message might contain the missing background, while Sentry might contain the relevant error logs. They wanted Claude, or another LLM, to reach into the outside world and bring back both context and actions. Chu describes this as a question of model agency. The model should be able to interact with external systems instead of receiving only the information manually pasted into its prompt. That led the team toward an open-source, standardized protocol that could work across tools and model providers.
An open standard removes a separate integration negotiation for every client
Chu argues that integrations become difficult when the client is part of a closed ecosystem. A builder may need business development or partnership access to the client team, then must agree on an interface before building the integration. An open protocol gives different participants a shared way to connect. Anthropic formed a small internal tiger team and launched MCP during its company hack week in November 2024, then open sourced it. Early reactions were skeptical. People asked why a new protocol was needed when models could already call tools. Adoption accelerated after Cursor and other coding tools added MCP, followed by adoption from Google, Microsoft, OpenAI, and others.
MCP is designed around agents choosing actions as model intelligence improves
Chu defines the agent direction around giving a model responsibility for choosing actions. A user can describe a task without specifying every step, and the model decides what to do based on its understanding of the request. That view affects protocol design. MCP added streamable HTTP as the primary transport instead of SSE, partly because it supports more bidirectional communication. Chu says this makes sense if agents need to communicate with one another. The team also expects many more servers than clients, although she admits that prediction could be wrong. As a result, MCP tends to put complexity on clients when it has to choose between client and server complexity, while keeping server building simpler.
Recent MCP work focuses on remote servers, authorization, and agent interaction
Chu lists several updates from the previous six months. MCP added support for remote servers and moved streamable HTTP into the primary transport position. The team corrected an initial mistake around OAuth in the draft specification after community members explained how identity providers work with the protocol. Anthropic also updated its SDKs and Inspector, which Chu calls an underused debugging tool. New work includes elicitation, which lets a server ask an end user for missing information. Her example is a flight booking request where the server needs to know whether 'best' means cheapest or fastest. The registry API should help models discover MCP servers that were not supplied in advance.
The strongest startup request is for more useful servers outside developer tools
Chu assigns 80 percent of her startup opportunity weighting to building more servers, compared with 10 percent for server-building tools and 10 percent for other areas. She argues that many current servers expose API endpoints one for one, which is not enough. A server has three users: the end user, the client developer, and the model. Builders need to understand likely user requests and prompts, then expose tools that let the model answer those requests correctly. Chu also wants servers for sales, finance, legal, education, and other fields. The ecosystem has many developer tools, but she considers the broader verticals largely open territory.
Server infrastructure can support both enterprise and independent builders
Chu expects the number of servers to exceed the number of clients by a large margin. That creates demand for tools that make server development easier for companies deploying MCP internally and for independent builders serving external users. She names hosting, testing, evaluation, and deployment as areas where more tooling is needed. She also describes automated MCP server generation as a longer-term bet. If models become capable enough at writing code and interacting with external systems, they may generate servers in real time for their own needs. Chu presents this as early relative to current model abilities, rather than as an immediately solved problem.
External model access increases the need for AI security and auditing
Chu's final startup category covers AI security, observability, auditing, and related tools. She says the need is broader than MCP because any AI application that can access external systems and real data creates similar concerns. MCP can give applications more reach into the outside world, so the consequences of weak security and privacy controls become more serious. Her recommendation is aimed at founders considering this space now. The talk ends with a direct connection between model agency and operational risk: as applications gain more ability to act, builders need ways to inspect, secure, and audit those actions.
"At some point models will be so good at writing code and interacting with the external world that they will actually be able to write their own MCPs on the fly in real time."16:38
Who should watch
You are deciding whether to build an MCP server and need guidance on tool design, target users, and industry opportunities.
You are building infrastructure for MCP developers and want areas such as hosting, testing, evaluation, or deployment to explore.
You work on AI security, observability, privacy, or auditing and want to understand why external model access creates demand for these products.
