Claude Code Changelog

Every Claude Code release, in order. Features shipped, bugs squashed, and anything that broke along the way. Coverage runs from the v0.2 beta up through March 2026.

Looking for the latest releases? This log stops at March 2026. Run claude --version in your terminal for whatever is current right now.

Version 2.1 Series (2026)

v2.1.81

• Bare mode. New --bare flag for scripted -p calls. Skips hooks, LSP, plugin sync, and skill directory walks. Needs ANTHROPIC_API_KEY or apiKeyHelper through --settings
• Channels permission relay. The --channels flag lets channel servers forward tool approval prompts to your phone
• OAuth session fix. Running sessions no longer force repeated logins when a sibling session refreshes its OAuth token
• Voice mode fixes. Retry failures stop getting swallowed silently, and audio recovers after the server drops the WebSocket
• Beta header fix. CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS now suppresses the structured-outputs beta header like it should

v2.1.80

• Channels feature. Plugin-based messaging for Telegram, Discord, and iMessage
• Bun runtime support. Plugins can run on Bun

v2.1.50

• Worktrees. First-class git worktree support behind a --worktree flag

Version 2.0 Series

v2.0.42

• SubagentStop fields. SubagentStop hooks now expose agent_id and agent_transcript_path

v2.0.41

• Stop hook model. Prompt-based stop hooks gain a model parameter
• Slash command fix. Commands from user settings stop loading twice
• Plugin hook fix. Crashes on plugin command hook timeouts are gone
• Bedrock model fix. Duplicate Opus entries no longer show up for Bedrock users

v2.0.37

• Notification idleness. Idleness math for notifications corrected
• Hook matchers. Notification hook events now accept matcher values
• Output styles option. Frontmatter gets a keep-coding-instructions switch

v2.0.35

• Fuzzy search. Command search quality improved
• VSCode fonts. The extension honors chat.fontSize and chat.fontFamily
• Exit delay. New CLAUDE_CODE_EXIT_AFTER_STOP_DELAY for scripted workflows

v2.0.34

• VSCode permission mode. New setting picks the starting permission mode for fresh chats
• Rust fuzzy finder. File path suggestions run on a native Rust fuzzy finder
• OAuth fix. MCP servers no longer get stuck in an OAuth token refresh loop
• Memory fix. Big files no longer crash on read or write

v2.0.32

• Output styles restored. Un-deprecated on community pushback
• Company announcements. New companyAnnouncements setting for startup messages

v2.0.31

• Windows mode switch. Native installs now use shift+tab to flip modes
• Vertex web search. Web search works on supported models
• VSCode gitignore. File search respects the new respectGitIgnore option

v2.0.30

• macOS keychain hint. API key errors now surface a security unlock-keychain tip
• Sandbox setting. New allowUnsandboxedCommands sandbox option
• Agent disallowedTools. Custom agents can list a disallowedTools field
• Prompt-based stop hooks. Stop hooks can now be prompt-based
• SSE MCP on native. Native builds support SSE MCP servers

v2.0.28

• Plan subagent. A new Plan subagent handles plan mode
• Subagent resume. Claude can resume a subagent on its own
• Dynamic model selection. Claude picks which model each subagent uses
• SDK budget flag. New --max-budget-usd flag

v2.0.27

• New permission UI. Permission prompts got a redesign
• Branch filtering. The session resume screen filters by current branch and supports search

v2.0.25

• Legacy SDK removed. The old SDK entrypoint is gone. Migrate to @anthropic-ai/claude-agent-sdk

v2.0.24

• Skills fix. Project-level skills now load under --setting-sources 'project'
• Web teleport. Claude Code Web supports Web -> CLI teleport
• Sandbox mode. BashTool sandbox mode shipped on Linux and Mac

v2.0.21

• MCP structuredContent. Tool responses support the MCP structuredContent field
• Interactive questions. A new interactive question tool for better back-and-forth
• Enhanced plan mode. Claude asks more clarifying questions while planning
• Haiku 4.5 for Pro. Pro users can pick Haiku 4.5

v2.0.20

• Claude Skills. Claude Skills are now supported

v2.0.19

• Auto-background commands. Long-running bash commands get sent to the background instead of killed
• Print mode fix. Haiku no longer gets called pointlessly in print mode

v2.0.17

• Haiku 4.5 released. Haiku 4.5 joins the model picker
• Smart model switching. Haiku 4.5 flips to Sonnet for planning and stays on Haiku to execute
• Explore subagent. A new Explore subagent runs on Haiku for cheap codebase searches

v2.0.12

• Plugin System Released. Extend Claude Code from marketplaces with custom commands, agents, hooks, and MCP servers
• Plugin commands. /plugin install, /plugin enable/disable, and /plugin marketplace for plugin management
• Team collaboration. Repo-level plugin config through extraKnownMarketplaces
• Plugin validation. /plugin validate checks plugin structure

v2.0.10

• Terminal renderer rewrite. The terminal renderer got rebuilt for a smoother UI
• MCP toggle. @-mention a server (or use /mcp) to flip it on or off
• Tab completion. Bash mode supports tab completion on shell commands
• Hook enhancements. PreToolUse hooks can now rewrite tool inputs
• Editor integration. Ctrl-G pops your prompt into the system editor

v2.0.0

• Native VS Code extension. Brand new native VS Code extension
• Fresh UI. A full visual refresh across the app
• Rewind feature. /rewind a conversation to roll back code changes
• Usage tracking. /usage shows your plan limits
• Thinking toggle. Tab toggles thinking, and the setting sticks across sessions
• History search. Ctrl-R searches history
• SDK rename. The Claude Code SDK is now the Claude Agent SDK
• Dynamic agents. Add subagents on the fly with --agents

Version 1.0 Series

v1.0.126

• Context command. /context now works on Bedrock and Vertex
• mTLS support. HTTP-based OpenTelemetry exporters support mTLS

v1.0.124

• Bash login control. Set CLAUDE_BASH_NO_LOGIN to skip the login shell for BashTool
• Security improvements. Claude no longer learns the allowed tool list when permission is denied
• Bash security. A permission-check hole in the Bash tool got patched

v1.0.123

• Bash permission rules. Output redirections work in rule matching
• Thinking mode fix. Phrases like "don't think" no longer trip thinking mode
• SlashCommand tool. A new SlashCommand tool lets Claude call your slash commands

v1.0.117

• History search. Ctrl-R brings a bash/zsh-style history search
• Input lag fix. Typing lag, especially on Windows, is fixed

v1.0.115

• Thinking display. Thinking mode gained richer visual effects
• Thinking toggle. Type /t in a prompt to pause thinking mode once

v1.0.110

• WezTerm support. /terminal-setup covers WezTerm
• OAuth refresh. MCP OAuth tokens refresh before they expire

v1.0.94

• Vertex global endpoints. Global endpoints now work for supported models
• Memory editing. /memory lets you edit imported memory files directly
• Todos command. A new /todos command lists active todo items

v1.0.90

• Settings reload. Settings changes take effect on save. No restart

v1.0.88

• OAuth fix. The "OAuth authentication is currently not supported" error is gone
• Status line. The status line now flags exceeds_200k_tokens
• Model environment variables. New ANTHROPIC_DEFAULT_SONNET_MODEL and ANTHROPIC_DEFAULT_OPUS_MODEL

v1.0.85

• Cost visibility. The status line shows live session costs
• Hooks enhancement. A new SessionEnd hook runs cleanup code

v1.0.81

• Output styles released. New "Explanatory" and "Learning" educational styles

v1.0.77

• Opus Plan Mode. A /model setting pins Opus to planning only

v1.0.73

• MCP flexibility. Multiple config files work: --mcp-config file1.json file2.json
• Alpine Linux support. musl-based distros are now supported

v1.0.71

• Background commands. Ctrl+B sends commands to the background, ideal for dev servers
• Custom status line. Build your own terminal status line with /statusline

v1.0.69

• Opus 4.1. A major model upgrade

v1.0.60

• Custom subagents. Build specialized subagents through /agents

v1.0.58

• PDF support. PDF reading added

v1.0.51

• Native Windows. Native Windows support landed. Requires Git for Windows

v1.0.38

• Hooks released. Custom automation hooks for pre-tool and post-tool events

v1.0.27

• MCP improvements. Streamable HTTP MCP servers, OAuth support, and @-mentionable resources
• Resume command. A /resume slash command to switch between conversations

v1.0.23

• SDK releases. TypeScript and Python SDKs shipped

v1.0.11

• Claude Pro support. Claude Code works under a Claude Pro subscription

v1.0.0

• General availability. Claude Code shipped GA
• New models. Sonnet 4 and Opus 4 introduced

Version 0.2 Series (Beta)

v0.2.96

• Claude Max. Claude Code works under a Claude Max subscription

v0.2.93

• Resume conversations. claude --continue and claude --resume
• Todo list. Claude gets a Todo list for organizing work

v0.2.75

• Message queuing. Hit Enter to stack extra messages while Claude is working
• Image support. Drag or paste image files straight in
• @-mentions. @-mention a file to push it into context

v0.2.54

• Memory shortcut. Prefix a message with # to add it to Memory
• Transcript mode. ctrl+r opens the full output for long tool results
• MCP SSE transport. SSE transport for MCP added

v0.2.50

• MCP project scope. Add MCP servers through .mcp.json and commit the file

v0.2.47

• Auto-complete. Tab auto-completes file and folder names
• Auto-accept toggle. Shift+Tab toggles auto-accept on file edits
• Infinite conversations. Conversations compact on their own

v0.2.44

• Thinking mode. Say "think", "think harder", or "ultrathink" to trigger planning

v0.2.34

• Vim bindings. Turn on via /vim or /config

v0.2.31

• Custom slash commands. Markdown files under .claude/commands/ show up as custom slash commands
• MCP debug mode. The --mcp-debug flag surfaces more info on MCP server errors

v0.2.21

• Fuzzy matching. Fuzzy matching lands for /commands