Detailed Analysis
A developer working on 10+ production projects with Claude Code as a daily driver on a 256GB SSD MacBook Air has documented in granular detail the hidden sources of disk consumption that silently degrade system performance on storage-constrained machines. The most significant offender identified is Claude Code's internal runtime cache stored in `~/Library/Application Support/Claude/vm_bundles/`, which the author found consuming 12GB at the time of writing and had previously observed exceeding 7GB before the folder's existence was even known. Additional contributors include per-session JSONL log files in `~/.claude/projects/` (each reaching 8–12MB with heavy tool use), an uncapped Google Drive for Desktop cache that had materialized 10GB of cloud-only files locally, and a 4GB Gemini Nano model silently downloaded by Chrome. Residual application data from fully uninstalled tools like Firefox, Opera, and Windsurf added another 3GB, highlighting how macOS's `~/Library/Application Support/` directory routinely retains orphaned caches long after parent applications are removed.
The broader technical picture corroborates what Anthropic's own GitHub issue tracker has been accumulating: Claude Code exhibits documented resource management problems beyond disk alone. Multiple open issues describe linear RAM growth during idle sessions, memory consumption reaching 8.5GB to 13GB RSS with 47GB of virtual memory on Windows, and process spawning behavior that degrades system performance over time. A particular feature called Cowork has been flagged for creating 10GB virtual machine bundles and allocating roughly 2GB of active memory even when the feature is never activated by the user. These issues appear connected to the desktop implementation's underlying architecture, with speculation pointing toward React-related inefficiencies in memory handling. As of April 2026, no official confirmed fixes have been documented in the issue threads, leaving users reliant on community-sourced workarounds.
The author's long-term mitigation strategy shifts from reactive cleanup to proactive monitoring, a distinction that carries practical weight. Rather than periodic manual audits, the author built an end-of-session debrief workflow that checks disk usage, RAM pressure, vm_bundles size, and session log growth, triggering an alert when disk utilization crosses 70%. Structural changes followed: 13 projects were moved off local storage entirely to GitHub and Google Drive, with local clones created on-demand and deleted after use. This repository-as-source-of-truth approach effectively decouples active development from persistent local storage footprint. The net result was approximately 70GB reclaimed, dropping overall disk usage from 81% to 58%, though the author notes it has since crept back to 74%, illustrating that the underlying growth mechanisms remain unaddressed at the tooling level.
The significance of this report extends beyond one developer's workflow optimization. Claude Code is increasingly positioned as a professional agentic coding tool, and its adoption among developers working on resource-constrained hardware — particularly MacBook Air configurations with 256GB storage, which represent a substantial share of the consumer MacBook market — means that silent resource consumption carries real consequences: failed builds, interrupted cloud sync, and degraded system responsiveness. The vm_bundles cache alone, if left unmonitored on a 256GB system, could consume nearly 5% of total available storage without any user action. The pattern also reflects a broader tension in AI desktop application development, where locally-deployed model components (like Chrome's Gemini Nano) and runtime caches are increasingly being installed without explicit user consent or visible resource accounting, a trend that will likely intensify as more AI features are bundled into mainstream applications.
The situation underscores a growing infrastructure challenge for Anthropic and similar companies shipping agentic AI tooling: the resource footprint of developer-facing AI tools has not kept pace with the expectation of lightweight, background-capable software. Developers running Claude Code in production contexts are effectively absorbing infrastructure costs — in storage, RAM, and monitoring overhead — that would typically be managed at the platform or SaaS layer. Community-driven workarounds like manual cache clearing scripts and context window compaction techniques are filling a gap that should be addressed through official memory management improvements, configurable cache size limits, and transparent resource usage dashboards. Until those mechanisms exist at the product level, storage hygiene on constrained machines remains, as the author puts it, not a nice-to-have but a survival requirement.
Read original article →