Claude docs changes for May 26th, 2026 [diff]

Executive Summary

  • Effort level is now part of the prompt cache key — switching it mid-session invalidates the cache and now triggers a confirmation dialog
  • New allowAllClaudeAiMcps managed setting lets enterprise admins load claude.ai connectors alongside a deployed managed-mcp.json
  • /usage now shows a usage breakdown by skill, subagent, plugin, and MCP server on Pro/Max/Team/Enterprise plans
  • Claude Security is now supported on GitHub Enterprise Server in public beta for Enterprise plans
  • New documentation clarifies which tools are unavailable to subagents in the Agent SDK (Agent, AskUserQuestion, EnterPlanMode, ScheduleWakeup, WaitForMcpServers)

Claude Code changes

Changed documents

agent-sdk/typescript [Source]

  • Critical clarification about the env option: it replaces the subprocess environment instead of merging with process.env, so callers must pass { ...process.env, YOUR_VAR: 'value' } to preserve inherited variables like PATH. [line 407] [Source]

agent-view [Source]

  • Clarified Ctrl+C behavior while attached to a session: it keeps its standard interrupt behavior (cancels a running response or ! shell command). Double Ctrl+C on an empty prompt detaches, same as in any session. [lines 153-154] [Source]

claude-directory [Source]

  • Added statsig/ and logs/ as additional legacy directories covered by the startup cleanup sweep (alongside the existing todos/). These are no longer written by current versions, and the sweep removes their contents and then the empty directories. [line 201] [Source]

commands [Source]

  • Updated /usage description to mention it now includes a breakdown of usage by skill, subagent, plugin, and MCP server on Pro, Max, Team, and Enterprise plans. [line 117] [Source]

costs [Source]

  • Added documentation for the new /usage breakdown feature: on Pro/Max/Team/Enterprise plans, shows usage attributed to skills, subagents, plugins, and individual MCP servers as a percentage of total. Press d or w to switch between last 24 hours and last 7 days. [lines 28-30] [Source]
  • Added note that Pro and Max subscribers can set a monthly spend limit on usage credits with /usage-credits, with a prompt to raise/remove the limit if it is reached mid-session. [line 34] [Source]

env-vars [Source]

  • Added new environment variable CLAUDE_CODE_SYNC_SKILLS: set to 1 to download enabled claude.ai skills into ~/.claude/skills/ before the first query and resync every 10 minutes. Applies only in non-interactive mode (-p). Set automatically in Claude Code on the web sessions. [line 242] [Source]
  • Added new environment variable CLAUDE_CODE_SYNC_SKILLS_WAIT_TIMEOUT_MS: timeout in milliseconds for the first query to wait on the initial skills sync (default: 5000). [line 243] [Source]

features-overview [Source]

  • Updated MCP description to clarify it provides "purpose-built tools for an external system, with connection and authentication handled by the server." [line 134] [Source]
  • Updated MCP reliability note: Claude Code now reconnects to remote servers automatically if they drop. [line 208] [Source]

github-enterprise-server [Source]

keybindings [Source]

  • Added pager-style keyboard shortcuts for the diff detail view: J/K for line-by-line scrolling, Space/Shift+Space and B for full-page scrolling, G/Shift+G and Home/End to jump to top/bottom. [lines 257-265] [Source]

managed-mcp [Source]

  • Clarified that deploying managed-mcp.json suppresses claude.ai connectors by default, including admin-configured organization connectors, unless the new allowAllClaudeAiMcps setting is enabled. [line 38] [Source]
  • Added new section documenting the allowAllClaudeAiMcps managed setting: set to true to load claude.ai connectors alongside managed-mcp.json. Requires v2.1.149 or later. Only effective from admin-controlled policy tiers; user/project settings have no effect. [lines 105-111] [Source]

mcp [Source]

  • Added troubleshooting guidance: if the plugin marketplace is not found when installing an MCP server plugin, run /plugin marketplace add anthropics/claude-plugins-official first before retrying the install. [line 42] [Source]

monitoring-usage [Source]

  • Clarified that the OTel headers helper value can be a path to an executable (including paths with spaces) or a shell command line with arguments. On Windows, the value always runs through the shell. [line 250] [Source]
  • Added documentation on where helper failures are reported: /doctor output, the debug log (with --debug), and stderr in non-interactive sessions. [lines 262-267] [Source]
  • Clarified that the tool_result OTel event is not emitted for rejected tool calls; decision_type on this event is always "accept". [lines 509-525] [Source]
  • Clarified user_reject vs config source distinction: in interactive CLI, deny-rule matches emit "config"; in Agent SDK and non-interactive sessions, deny-rule matches emit "user_reject". [line 629] [Source]

output-styles [Source]

  • Added note that the /output-style command was deprecated in v2.1.73 and removed in v2.1.91; use /config or edit the outputStyle setting directly. [line 25] [Source]

permissions [Source]

  • Clarified that the permissions.additionalDirectories setting in a settings file grants file access only and does not trigger configuration discovery. The --add-dir flag and /add-dir command are the only paths that load configuration from additional directories. [line 260] [Source]
  • Added allowAllClaudeAiMcps to the managed-settings-only settings table. [line 301] [Source]

prompt-caching [Source]

  • Effort level is now documented as part of the cache key: each effort level has its own cache for the same model. Switching mid-session recomputes the entire request. [lines 28-29] [Source]
  • Claude Code now shows a confirmation dialog before applying an effort level change that would invalidate the cache. A change that resolves to the same level already in effect skips the dialog. [lines 62-65] [Source]

remote-control [Source]

  • Added note that renaming a session from claude.ai or the Claude app also updates the local title shown in claude --resume. [line 109] [Source]

settings [Source]

  • Added allowAllClaudeAiMcps to the settings reference: managed-settings-only option to load claude.ai connectors alongside managed-mcp.json. [line 158] [Source]

skills [Source]

  • Clarified that permissions.additionalDirectories in settings.json does not load skills; only the --add-dir flag and /add-dir command trigger skill loading from additional directories. [line 129] [Source]
  • Added new section "How a skill gets its command name" explaining that the invocation name comes from the skill's file/directory location, not the frontmatter name field (except for plugin-root SKILL.md files where name sets the command name). [lines 193-207] [Source]

sub-agents [Source]

  • Added documentation for tools that are unavailable to subagents even when listed in the tools field: Agent, AskUserQuestion, EnterPlanMode, ExitPlanMode (unless the subagent's permissionMode is plan), ScheduleWakeup, and WaitForMcpServers. [lines 291-298] [Source]

tools-reference [Source]

  • Clarified that EnterWorktree and ExitWorktree are not available to subagents that already run in their own working directory (e.g., with isolation: worktree). [lines 22-23] [Source]

API changes

Changed documents

adaptive-thinking [Source]

  • Updated the contact link for requesting full thinking output access for Claude 4 models to use a direct mailto:sales@anthropic.com address instead of an obfuscated link. [line 179] [Source]