v2.1.136

Source: Anthropic Claude Code Date: 2026-05-08 URL: https://github.com/anthropics/claude-code/releases/tag/v2.1.136

Summary

Claude Code v2.1.136 is a large quality-of-life and stability release with 50+ fixes. The most operationally significant additions are settings.autoMode.hard_deny (unconditional classifier blocks that override user intent), a fix for MCP servers silently vanishing after /clear in VS Code/JetBrains/Agent SDK, and resolution of OAuth token-race conditions that caused daily re-authentication across multiple MCP servers. A pile of terminal rendering, plugin, and slash-command regressions are also closed out.

Implications

• hard_deny in auto mode is the first explicit hook for enterprise policy teams to enforce unconditional agent blocks — signals Anthropic is building out governance primitives for fleet deployments, not just individual developer guardrails.
• The MCP persistence fixes are directly load-bearing for multi-server agentic setups; silent disappearance after /clear was a reliability gap that would have surfaced in any long-running orchestration session.
• Volume and character of fixes (50+ in one release) suggests rapid iteration velocity; the surface area of edge cases around MCP OAuth, plugin hooks, and terminal rendering is growing with adoption.