Claude docs changes for June 20th, 2026 [diff]

Executive Summary

New disableClaudeAiConnectors setting lets you disable claude.ai MCP connectors at any settings scope (user, project, org), giving per-project control over cloud connectors
Auto permission mode expanded to block more destructive operations: git reset --hard, git stash drop/clear, git commit --amend on pre-session commits, and Terraform/Pulumi/CDK destroy commands
New SDKRateLimitEvent fields (errorCode, canUserPurchaseCredits, hasChargeableSavedPaymentMethod) allow SDK callers to detect and handle credits-exhausted subscription scenarios
/config command now accepts named shorthand keys (e.g. /config theme=dark, /config model=sonnet) and --help lists all settable keys
Two new env vars: CLAUDE_CODE_CONNECT_TIMEOUT_MS for API streaming timeouts and CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS for capping background subagent wait in non-interactive mode

No new versions today.

Added disableClaudeAiConnectors: true as an alternative to strictMcpConfig: true or the ENABLE_CLAUDEAI_MCP_SERVERS=false env var for suppressing claude.ai MCP connectors. [line 57] [Source]

Added three new optional fields to SDKRateLimitEvent: errorCode (set to "credits_required" when a claude.ai subscription's included usage is exhausted), canUserPurchaseCredits, and hasChargeableSavedPaymentMethod. Requires Claude Code v2.1.181 or later. [lines 3234-3248] [Source]

From v2.1.182, the new attribution.sessionUrl setting can be set to false to omit the Claude-Session git trailer and PR-body session link that web sessions add by default. [line 104] [Source]

/config now accepts named shorthand keys as of v2.1.182, such as /config theme=dark or /config model=sonnet. The help flag changed from help to --help, and the description now says it lists every settable key with its options. [line 46] [Source]

New CLAUDE_CODE_CONNECT_TIMEOUT_MS env var sets the timeout (default 60s) for the connect, TLS, and response-header phase of a streaming API request. Set to 0 to rely on API_TIMEOUT_MS alone. [line 149] [Source]
New CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS env var controls how long non-interactive mode (-p) waits for background subagents and workflows. Default is 10 minutes; set to 0 to wait indefinitely. [line 230] [Source]
ENABLE_CLAUDEAI_MCP_SERVERS documentation now notes that disableClaudeAiConnectors in settings is the preferred way to disable connectors per-project or per-org. [line 303] [Source]

Background subagents and workflows are now exempt from the 5-second grace period and instead wait until complete. From v2.1.182, that wait is capped at 10 minutes by default to prevent a stuck agent from holding the process open indefinitely. Configurable via CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS. [lines 52-53] [Source]

Settings context actions updated: settings:close replaced by select:accept (Enter/Space) and confirm:no (Escape). Clarification added that settings apply immediately as you change them, so Escape closes with changes saved. [lines 296-302] [Source]

Expanded explanation of serverName validation differences between allowedMcpServers and deniedMcpServers: in deny lists, any non-empty string is accepted (including claude.ai connector display names like "claude.ai Slack"); in allow lists, only letters, numbers, hyphens, and underscores are accepted. [lines 129-137] [Source]

New "Disable claude.ai connectors" section documents the disableClaudeAiConnectors settings key (any-source-true semantics), the existing ENABLE_CLAUDEAI_MCP_SERVERS=false env var alternative, and guidance on blocking individual connectors via deniedMcpServers. Also clarifies that disableClaudeAiConnectors does not apply to Claude Code on the web sessions. [lines 742-769] [Source]

From v2.1.182, model retirement and remapping warnings are now written to stderr in non-interactive mode when using the default text output format. The warning is suppressed for --output-format json and stream-json; use the modelUsage field of the result message instead. [line 69] [Source]

Auto permission mode now blocks additional destructive operations: git reset --hard, git checkout -- ., git restore ., git clean -fd, git stash drop, git stash clear (presumed to discard uncommitted changes); git commit --amend when the HEAD commit was not created in this session; and terraform destroy, pulumi destroy, cdk destroy, terragrunt destroy (plus any plan that destroys resources). [lines 180-182] [Source]

New disableClaudeAiConnectors setting documented: disables claude.ai MCP connectors globally. Uses any-source-true semantics — a project-level false cannot override a user- or policy-level true. Requires v2.1.182+. [line 214] [Source]
New remoteControlAtStartup setting: connects Remote Control automatically at every interactive session start. Appears in /config as "Enable Remote Control for all sessions". [line 257] [Source]
New attribution.sessionUrl sub-setting: controls whether the claude.ai session link is appended as a Claude-Session trailer on commits and a link in PR descriptions. Defaults to true. [line 423] [Source]

New "Set the response language" section added with code examples and guidance on explicitly setting the response language in the system prompt for production applications. [lines 43-57] [Source]
Expanded "Next steps" section with links to customer support agent guide, models overview, and evaluation guidance. [lines 82-93] [Source]