AI labs keep moving into the layer their customers build on
The top comment on 'Sites in ChatGPT' asks OpenAI to just be an API instead of moving into the app layer and putting clients out of business. Another commenter notes it closes a gap that both Claude and ChatGPT had: asking for a small website and actually getting one. Separately, Anthropic stopped letting Claude Code subscriptions power OpenClaw, citing 'outsized strain', and offered a one-time credit. Users answered that this is just using the capacity they paid for.
The pattern is the same from two angles. The labs are absorbing features that third-party tools used to supply, and they are tightening what their flat-rate plans can be used for. Offrun, a Show HN for managing every coding agent from one workspace, got the first question every thin layer gets: what does this do that Herdr doesn't? A second commenter said an agent had pitched them nearly the same product that same week.
Some builders read this as a sign that tools should be agent-agnostic and portable across CLIs. One commenter is experimenting with MCP plus a custom UI for exactly that reason.
So what?
Orchestration and wrapper products are crowded and sit on someone else's pricing rules. If your product is a thin layer over one lab's subscription, assume that lab will either ship your feature or cut your access. The defensible ground is proprietary data, workflow depth and being model-agnostic.