MCP Servers Become Distribution Deadline
Reed McGinley-Stempel, CEO of Stytch, on authentication for AI agents
This turns MCP into a distribution deadline, not an infrastructure side project. For companies that used to depend on Google search and their own website, an MCP server is becoming the way their product and content show up inside Claude, ChatGPT, and coding tools. That pushes identity work up to the CEO level because shipping late can mean a competitor becomes the default tool an agent can reach first.
-
The practical change is that apps must act more like Google or Microsoft do for third party access. They need OAuth based login, consent screens, scoped permissions, revocation, and audit logs so an agent can read data or take actions without handing out raw API keys.
-
The first MCP wave was read only access to paid data inside AI clients, then write actions started following. Examples include biotech content delivered through MCP, Linear style actions through Claude, and integration vendors like Ampersand and Zapier launching MCP products to become the connection layer for agent workflows.
-
The urgency is tied to discovery. Anthropic already supports remote MCP servers across Claude products and its API, including OAuth for authenticated servers. OpenAI is also training users to expect connected apps inside ChatGPT through connectors, which makes being absent from these surfaces look increasingly like being absent from search results.
The next step is a land grab for agent accessible surfaces. More companies will rush to expose narrow, high value workflows first, then expand into deeper read and write permissions. The winners will be the products that become easy for agents to reach, safe for enterprises to approve, and useful enough to become the default path instead of the website.