v2.1.187

Source: Anthropic Claude Code Date: 2026-06-23 URL: https://github.com/anthropics/claude-code/releases/tag/v2.1.187

Summary

Claude Code v2.1.187 is a substantial patch release focused on agent reliability and SDK hardening. Key additions include a sandbox.credentials setting to prevent sandboxed commands from accessing secrets, org-level model restrictions surfaced throughout the UI, and mouse-click support in fullscreen select menus. The release closes a long tail of agent-loop bugs: hung remote MCP tool calls (now abort after idle timeout), background jobs stuck in “working,” subagent depth tracking on resume/fork, and leaked worktree registrations left behind by killed agents.

Implications

• Dev-tooling floor: The credential-isolation and MCP idle-timeout fixes are the kind of hardening that moves Claude Code from “powerful but sharp-edged” toward production-safe. Each release narrows the gap between what developers can demo and what they can actually ship in a org context.
• Agent landscape: Subagent depth tracking and forked-agent depth caps signal that multi-agent topologies are a first-class concern, not an afterthought — the runtime is being engineered for trees of agents, not just single-session loops.
• Governance/safety: Org-configured model restrictions in the model picker close a quiet compliance gap; teams on enterprise contracts can now enforce which models run without relying on honor-system prompting.
• Voices/power dynamics: The optional GitHub App workflow setup in /install-github-app lowers the on-ramp for teams that want the app integration without a full CI rewrite — friction removal that tends to expand the install base.