Detailed Analysis
The Claude Mac application's "Cowork" feature includes an @ mention functionality designed to allow users to reference open application windows, providing Claude with contextual visual information about what is displayed on screen. The feature appears conceptually similar to OpenAI Codex's "appshot" capability, which captures and processes visual content from active applications to give an AI assistant real-time awareness of the user's working environment. However, users attempting to employ this feature are encountering a blocking error message stating that content "was uploaded for a remote session and can't be sent to a local one," with instructions to either remove and re-attach the content or remove a working folder to stay remote — even when no working folder has been explicitly configured by the user.
The core technical issue appears to stem from a session-type conflict within the Cowork architecture. The Claude Mac app seemingly distinguishes between "local" and "remote" processing sessions, and the @ window operation is triggering an upload pathway associated with remote sessions, while the active session is classified as local. This mismatch prevents the window content from being delivered to Claude as intended, rendering the feature non-functional for affected users. The fact that this occurs without the user having set a working folder suggests that either the application is automatically inferring session type in unexpected ways, or there is a bug in how session state is being managed and communicated to the user.
This user experience report highlights a broader challenge in shipping polished desktop AI integration features — specifically the complexity of managing context-sharing mechanisms across different execution environments. The ambiguity around what constitutes a "remote" versus "local" session, combined with error messaging that presupposes user familiarity with internal session architecture, points to gaps in both feature stability and user-facing documentation at the time of this post.
The broader significance lies in the competitive landscape around AI desktop assistants gaining ambient awareness of user workflows. Features that allow AI tools to observe and respond to on-screen application content represent a meaningful capability frontier, one that both Anthropic and OpenAI are actively developing. When these features malfunction without clear resolution paths, user trust erodes and adoption of more sophisticated agentic workflows is slowed. The Cowork @ window issue underscores the difficulty of delivering seamless, cross-context AI assistance in desktop environments where session management, local compute, and cloud processing intersect in ways users are not expected to understand or manually resolve.
Read original article →