Anthropic appears to be testing “@” mentions inside the Claude Cowork chat composer. Typing or tapping “@” would open a pop-up of callable actions from connected MCPs. One example is a Puppeteer MCP action that can pull browser console logs into the chat, suggesting that Cowork is being tuned for tighter tooling and debugging workflows inside a single thread.
What stands out is that the same @ menu also surfaced sample entries that do not read like MCP tools but like app-style actions: “hello world,” “counter,” plus “imagine SVG widget” and “imagine HTML widget.” This mirrors the ChatGPT apps concept, where apps can return richer, structured outputs into the chat UI, including embedded widgets rather than plain text. If Claude adopts a comparable approach, it could become a second major surface for developers building chat-native apps that render UI in-line, potentially across multiple platforms with similar patterns.
Separately, earlier “knowledge base” traces seem to have shifted toward project selection, implying projects may be the intended container for persistent context as their memory updates over time. There are also signs of a revamped chat search, with a dedicated pop-up for querying past conversations, aimed at making long-term usage easier to navigate.
