Detailed Analysis
A Reddit post highlights user confusion over a recent, apparently unannounced change to Anthropic's Claude desktop client, specifically the Mac app. According to the poster, Claude Code (referred to informally as "cowork" sessions) had previously defaulted to launching local, on-device Claude instances. As of the morning the post was written, that default silently switched to cloud-based Claude sessions instead. The user reports spending roughly thirty minutes trying to locate the toggle to switch back to local execution, eventually finding it hidden behind what they describe as an unlabeled, icon-only button with no clear indication of its function.
This complaint sits at the intersection of two ongoing threads in Anthropic's product rollout: the introduction of Claude Code's cloud execution environment a few weeks prior, and the friction that comes when a company changes default behaviors without clear communication or in-app signaling. Local versus cloud execution is not a trivial distinction for developers using Claude Code — local runs execute directly against a user's file system and environment, while cloud runs operate in a remote sandboxed environment, which affects file access, latency, persistent state, and how a session integrates with local tooling, git repositories, and terminal contexts. A silent default switch from local to cloud could mean sessions no longer have the same access to a developer's local machine unless manually reconfigured, which is a meaningful workflow disruption for anyone relying on Claude Code as a coding assistant embedded in their local dev environment.
The underlying frustration reflects a broader pattern seen across the AI coding-assistant space: rapid, iterative shipping of features and default changes, often without robust changelogs, in-app tooltips, or migration notices reaching the average user. Anthropic has been aggressively expanding Claude Code's capabilities — adding cloud execution, sandboxing, and various orchestration options — as it competes with tools like GitHub Copilot Workspace, Cursor, and OpenAI's Codex-based offerings. In that competitive race to ship features quickly, UX polish and discoverability sometimes lag behind functional capability, and toggles or settings can get buried in menus that assume users are already tracking every product update via release notes or social channels rather than through the interface itself.
More broadly, this kind of anecdote is a useful signal for how power users experience the pace of AI tool development in 2026: capability is advancing quickly, but the operational scaffolding around that capability — clear defaults, discoverable settings, migration warnings — often struggles to keep pace. For a tool like Claude Code, where the distinction between "runs on my machine" and "runs in Anthropic's cloud" carries real implications for privacy, file access, and workflow continuity, defaulting users into a new execution mode without clear notice is the kind of decision that generates exactly this sort of Reddit-thread friction. It also underscores a recurring theme in AI product development broadly: as these tools become more agentic and infrastructure-heavy (spinning up cloud sandboxes, managing remote execution), the settings surface required to control them grows more complex, and companies are still working out how to make that complexity legible to everyday users rather than just technically accessible.
Read original article →