Claude docs changes for May 22nd, 2026 [diff]

Executive Summary

• Auto mode is now available to all users on the Anthropic API (previously restricted to Max, Team, Enterprise, and API plans)
• /simplify has been renamed to /code-review with new behavior: it now reports correctness bugs at a configurable effort level and can post inline GitHub PR comments with --comment
• Two new documents added: a dedicated guide for organizational MCP server control (managed-mcp.md) and a new sandbox environments comparison guide (sandbox-environments.md)
• MCP servers now support a per-server timeout field in .mcp.json, overriding the global MCP_TOOL_TIMEOUT for individual servers
• Git for Windows is now optional on native Windows installs — Claude Code falls back to the PowerShell tool when Git Bash is absent

New Claude Code versions

2.1.147

New features

• Pinned background sessions (Ctrl+T in claude agents) now stay alive when idle, are restarted in place to apply Claude Code updates, and are shed under memory pressure only after non-pinned sessions
• Renamed /simplify to /code-review — it now reports correctness bugs at a chosen effort level (e.g., /code-review high); pass --comment to post findings as inline GitHub PR comments. The old cleanup-and-fix behavior has been removed

Existing feature improvements

• Improved auto-updater: retries transient network failures, reports specific error categories and OS error codes on failure, and shows the current version when an update fails
• Improved diff rendering performance for large file edits
• Prompt history no longer records consecutive duplicate entries — recalling a prompt with arrow-up and submitting it again won't add another copy

Major bug fixes

• Fixed enterprise login restrictions (forceLoginOrgUUID and forceLoginMethod managed-settings) not being enforced against third-party-provider and API-key sessions
• Fixed & in ! command output displaying as &, which broke copy-pasting URLs from commands like gcloud auth login on headless machines
• Fixed unknown slash commands silently doing nothing in headless/SDK mode — they now show an error message
• Fixed plugin agents that declare multiple Agent(...) types in tools: frontmatter dropping all but the last entry
• Fixed hook if conditions like PowerShell(git push*) never matching — only PowerShell(*) worked
• Fixed PowerShell tool failing on Windows with exit code 1 when pwsh is installed via winget or the Microsoft Store
• Fixed paginating MCP servers dropping resources, templates, and prompts past page 1
• Fixed auto mode suppressing AskUserQuestion when the user or a skill explicitly relies on it
• Fixed pasted text being delivered to agents as an unreadable [Pasted text #N] placeholder instead of the actual content
• Fixed backgrounded sessions re-prompting for tool permissions you already granted with "don't ask again"
• Fixed CLAUDE_CODE_SUBAGENT_MODEL not applying to teammate processes spawned by agent teams
• Fixed slash commands followed by a tab or newline being treated as an unknown command
• Fixed an uncaught exception at the end of streaming sessions when running via the Agent SDK

2.1.148

Major bug fixes

• Fixed the Bash tool returning exit code 127 on every command for some users (a regression introduced in 2.1.147)

Claude Code changes

New Documents

Control MCP server access for your organization [Source]

New dedicated guide for administrators controlling MCP server access in an organization. Covers deploying a fixed server set with managed-mcp.json (including how to disable MCP entirely), restricting servers with allowedMcpServers/deniedMcpServers allowlists and denylists, how a server is evaluated through the full chain of checks, how restrictions appear to users, and monitoring MCP usage via OpenTelemetry. The managed MCP configuration content was previously embedded in mcp.md and has been moved and significantly expanded here.

Choose a sandbox environment [Source]

New comparison guide covering all available isolation approaches: the sandboxed Bash tool, sandbox runtime (@anthropic-ai/sandbox-runtime), dev containers, custom containers, virtual machines, and Claude Code on the web. Explains how isolation relates to permission modes (including --dangerously-skip-permissions and auto mode), and covers how organizations can enforce isolation across all developers.

Changed documents

Admin setup [Source]

• MCP server control row updated to reference the new managed-mcp.md guide and now explicitly mentions deploying a fixed server set via managed-mcp.json [line 66] [Source]

Agent SDK TypeScript reference [Source]

• Added clarification on the signal field in SpawnOptions: the signal does not fire the instant abortController aborts — the SDK first closes stdin and waits ~2 seconds for clean shutdown before aborting [line 2864] [Source]

Agent view [Source]

• Pinned sessions (Ctrl+T) now keep their process running while idle, and idle pinned sessions are restarted in place onto new Claude Code versions [line 382] [Source]
• When the host runs low on memory, the supervisor stops idle non-pinned sessions first [line 384] [Source]

Auto mode configuration [Source]

• Auto mode is now available to all users on the Anthropic API (previously Max, Team, Enterprise, and API plans only) [line 11] [Source]

Commands [Source]

• Added /code-review command entry with full description: reviews the diff for correctness bugs at a configurable effort level, supports --comment to post inline GitHub PR comments, formerly /simplify [line 42] [Source]
• Removed /simplify entry (replaced by /code-review)
• Updated "Before you ship" workflow tip to reference /code-review instead of /simplify

Desktop [Source]

• Auto mode plan requirements updated: now available to all users on the Anthropic API, not just Max/Team/Enterprise/API plans [line 74] [Source]

Dev container [Source]

• Added link to new sandbox-environments.md in the "See also" section [line 178] [Source]

Environment variables [Source]

• MCP_TOOL_TIMEOUT description now notes that a per-server timeout field in .mcp.json overrides it for that server, and values below 1000 are floored to one second [line 306] [Source]
• CLAUDE_CODE_ALT_SCREEN_FULL_REPAINT description updated to note it is also auto-enabled for agent view on Windows (not just background sessions) [line 136] [Source]

Errors [Source]

• Added new "Unable to resize image" error section covering four variants of the error with specific fix guidance [lines 465-483] [Source]
• 5xx error messages now reflect the inference provider (Bedrock, Vertex AI, Foundry, or custom gateway) rather than being Anthropic-specific [lines 64-100] [Source]
• 429 rate limit error: trailing sentence now also names Foundry and custom ANTHROPIC_BASE_URL gateways [line 201] [Source]

Glossary [Source]

• Auto mode entry updated: now available to all users on the Anthropic API [line 40] [Source]
• Bundled skills entry updated to reference /code-review instead of /simplify [line 52] [Source]

Interactive mode [Source]

• Added note that submitting the same prompt twice records only one history entry [line 208] [Source]

Keybindings [Source]

• permission:toggleDebug is now unbound — the previous default of Ctrl+D was removed in v2.1.146 because it shadowed app:exit [line 152] [Source]

MCP [Source]

• Added per-server timeout field: set "timeout": <ms> in a server's .mcp.json entry to override MCP_TOOL_TIMEOUT for that server only. Values below 1000 are floored to 1 second; HTTP/SSE first-byte budget has a 60-second minimum [lines 157-159] [Source]
• The large "Managed MCP configuration" section has been replaced with a short pointer to the new managed-mcp.md guide [line 996] [Source]

Model configuration [Source]

• CLAUDE_CODE_SUBAGENT_MODEL now applies to agent teams in addition to subagents, and supports inherit to use normal model resolution [line 273] [Source]

Permission modes [Source]

• Auto mode plan requirement updated: now available on all plans; supported models list no longer distinguishes Max vs. other plans [lines 141-144] [Source]
• Auto mode now notes it still allows AskUserQuestion when a prompt or skill explicitly relies on it [line 141] [Source]

Permissions [Source]

• allowManagedPermissionRulesOnly clarified: it does not affect the MCP server allowlist; use allowManagedMcpServersOnly for that [line 302] [Source]

Plugin marketplaces [Source]

• Three new reserved marketplace names added: claude-for-legal, claude-for-financial-services, financial-services-plugins [line 171] [Source]

Sandboxing [Source]

• Major restructure: renamed to "Configure the sandboxed Bash tool" with new step-by-step setup guide (/sandbox → choose mode → run a command) [lines 18-50] [Source]
• Ubuntu 24.04 AppArmor check now includes a command to verify whether the restriction applies before applying the profile fix [line 85] [Source]
• Added new "How sandboxing relates to permissions and permission modes" section with a clear table comparing sandbox settings vs. permission rules [lines 165-210] [Source]
• Default read behavior now explicitly warns that credential files (~/.aws/credentials, ~/.ssh/) are readable by default and should be added to denyRead to block them [line 172] [Source]
• allowUnsandboxedCommands: false ("Strict sandbox mode") now shown in the /sandbox Overrides tab [line 133] [Source]

Scheduled tasks [Source]

• Removed the statement that the built-in maintenance prompt "isn't available to everyone yet"; it now simply states it is unavailable on Bedrock, Vertex AI, and Foundry [line 85] [Source]

Security [Source]

• Added link to new sandbox-environments.md in related resources [line 132] [Source]

Server-managed settings [Source]

• Limitation note updated: clarifies that managed-mcp.json cannot be distributed through server-managed settings, but allowedMcpServers and deniedMcpServers policy keys can be [line 123] [Source]

Settings [Source]

• forceLoginMethod now documented to block API key, apiKeyHelper, and third-party provider sessions when set in managed settings [line 201] [Source]
• forceLoginOrgUUID similarly clarified to block API key and third-party provider sessions (organization membership cannot be verified for them) [line 202] [Source]

Setup [Source]

• Git for Windows is now optional on native Windows — Claude Code uses the PowerShell tool when Git Bash is absent [lines 91-117] [Source]
• Windows setup section renamed from "Native Windows with Git Bash" to "Native Windows" [line 91] [Source]

Skills [Source]

• Updated bundled skills list to reference /code-review instead of /simplify [lines 12-16] [Source]

Terminal guide [Source]

• Git for Windows install step now marked as optional; updated description clarifies it enables the Bash tool, and Claude Code falls back to PowerShell without it [lines 65-70] [Source]
• Error message fix for "Claude Code on Windows requires either Git for Windows or PowerShell": guidance reordered to lead with ensuring powershell.exe is on PATH [lines 241-250] [Source]

Tools reference [Source]

• PowerShell tool now documented to spawn with -ExecutionPolicy Bypass at process scope; does not override Group Policy. Set CLAUDE_CODE_POWERSHELL_RESPECT_EXECUTION_POLICY=1 to use the machine's effective policy instead [line 211] [Source]

Troubleshoot install [Source]

• Added curl: (22) 403 error to the quick-reference table and expanded the install-script-returns-HTML section to explain bare 403 responses (can be a proxy/firewall issue, not just regional blocking) [lines 18-20, 272-279]
• curl: (56) section now also covers curl: (23) and explains the difference between the two exit codes [line 313] [Source]
• Windows shell error guidance updated: Git for Windows is now optional; fix leads with adding powershell.exe to PATH [lines 494-504] [Source]

API changes

Changed documents

Deploy MCP tunnels with Docker Compose [Source]

• Container and config file renamed from mcp-gateway to mcp-proxy throughout: service name, volume mount (mcp-gateway.yaml → mcp-proxy.yaml), and network_mode reference [lines 103-130, 194]

MCP tunnels quickstart [Source]

• Docker Compose service and config file renamed from mcp-gateway to mcp-proxy in all quickstart examples and log commands [lines 123-177] [Source]

MCP tunnels reference [Source]

• Default --output Kubernetes secret name changed from k8s-secret:mcp-gateway to k8s-secret:mcp-tunnel for both the setup and renew-cert subcommands [lines 81, 93]

MCP tunnels troubleshooting [Source]

• Troubleshooting table and restart commands updated to use mcp-proxy service name instead of mcp-gateway [lines 14-15, 35, 109]