2026-04-28 · Anthropic

v2.1.121

securityprotocolsagentsmodels

read at source ↗ github.com

v2.1.121

Source: Anthropic Claude Code Date: 2026-04-28 URL: https://github.com/anthropics/claude-code/releases/tag/v2.1.121

Summary

Claude Code v2.1.121 shipped a broad maintenance release covering 40+ changes. Notable additions include an alwaysLoad MCP server option that bypasses tool-search deferral, claude plugin prune for orphaned dependency cleanup, PostToolUse hooks that can now replace tool output for any tool (not just MCP), and fork-subagent support in non-interactive SDK sessions. Memory leak fixes addressed multi-GB RSS growth from image-heavy sessions and large transcript histories.

Implications

  • Claude Code / agentic tooling thread: The alwaysLoad MCP option and broader PostToolUse hook coverage reduce friction in multi-server agent setups, making it easier to build pipelines where tool outputs are intercepted and rewritten programmatically.
  • Developer ergonomics: Fullscreen scroll stability, scrollable overflow dialogs, and MCP auto-retry on transient errors address real pain points in long-running agentic sessions.
  • SDK/automation thread: CLAUDE_CODE_FORK_SUBAGENT=1 in non-interactive mode and mcp_authenticate redirect URI support expand the surface area for headless orchestration patterns.

← all signals