Anthropic Claude Code v2.1.128: 30+ Fixes, .zip Plugin Support and ~3× Lower cache_creation Cost for Sub-Agents

Anthropic released Claude Code v2.1.128 on May 4, 2026 — a release that brings more than 30 improvements and fixes across the entire CLI stack. While there are no new flagship features, the combination of performance optimizations, plugin support and bug fixes makes this release practically important for teams that make heavy use of sub-agents and the custom plugin ecosystem.

What does the new MCP, plugin and agent layer bring?

The most impactful change is in the sub-agent layer: cache_creation cost — i.e., the cost of initially creating the prompt cache — has been reduced by approximately threefold. This was achieved by embedding prompt cache data into sub-agent progress summaries, allowing each new sub-agent invocation to reuse existing cache hits instead of creating a new cache layer every time. For workflows with parallel sub-agents (scout/writer agent orchestration like that used in our pipeline) this means a direct reduction in API costs.

The /mcp panel now displays the number of tools per connected server and explicitly flags servers that connected but returned no tools — a common silent mode failure (the server works but the configuration doesn’t expose tools). Additionally, the name workspace has been declared reserved — existing servers with that name will be skipped with a warning. The reconnect optimization consolidates re-announced tools by server prefix instead of flooding the conversation with full lists.

The plugin system gains support for .zip archives in the --plugin-dir flag — distributing plugins across a team no longer requires syncing a directory structure via git/rsync. Also fixed: /plugin update that failed to detect new versions of npm-sourced plugins, stale installed_plugins.json entries that polluted PATH, and the Components panel showing “Marketplace ‘inline’ not found” for --plugin-dir plugins.

What are the key bug fixes?

EnterWorktree no longer loses local unpushed commits — the new branch is now created from the local HEAD as the documentation always stated, rather than from origin/<default-branch>. The crash loop when piping inputs larger than 10 MB to claude -p via stdin is fixed. Drag-and-drop image upload no longer hangs on “Pasting text…” when image read fails. Parallel shell calls: a failing read-only command (grep, git diff, ls) no longer cancels sibling calls.

Minor but useful fixes: 1M-context models with a smaller autocompact window no longer receive a false “Prompt is too long” before the API limit; OTEL_* environment variables no longer leak into subprocesses (Bash, hooks, MCP, LSP), which was causing OTEL-instrumented applications to pick up the CLI’s OTLP endpoint; vim NORMAL mode Space now moves the cursor right (per vi/vim standard).

What does this mean for enterprise use?

The release confirms sub-agent economics and plugin distribution as the primary axes of Claude Code development. The cache_creation cost reduction directly impacts enterprise teams already measuring sub-agent token utilization in CI/CD pipelines. .zip plugin distribution simplifies compliance review (one artifact instead of a directory tree). The consistent fixing of session-state and worktree bugs signals that Anthropic is hardening the CLI for long-running enterprise workspaces, where any commit loss from a worktree workflow can mean the loss of significant work.