Claude Code Changelog: What's New, Version by Version

The Claude Code changelog is the running record of every Claude Code release, and the biggest changes in 2026 are Claude Sonnet 5 becoming the default model with a 1M-token context window (v2.1.197, June 30, 2026), subagents running in the background by default (v2.1.198, July 1, 2026), and agent teams collapsing into a single implicit team per session (v2.1.178, June 15, 2026). This page is a curated, community-maintained changelog: it summarizes the notable entries with what changed and why it matters, and it is refreshed as new builds ship. For the exhaustive per-build list, always check the official Claude Code changelog and the CHANGELOG.md on GitHub.

How to Read This Changelog

This changelog is reverse-chronological: the most recent release sits at the top, and each entry gives the version number, the date drawn from the official changelog, what changed, and why it matters. Claude Code ships small builds almost daily, so this page groups the notable ones rather than reprinting every patch line. The canonical source is Anthropic's official changelog; this page exists to add plain-language context and links to deeper guides.

If you are new to the tool, start with our complete guide to Claude Code and the installation walkthrough, then come back here to track what changes over time.

Recent Releases (2026)

Claude Code 2.1.201 is the most recent build referenced in the official changelog as of July 2026, and the 2.1 line has been the platform for the model, subagent, and agent-team changes below. Entries run newest first.

v2.1.198 to v2.1.201 - early July 2026

What changed: Subagents now run in the background by default, Claude in Chrome reached general availability, and a built-in /dataviz skill was added for chart and dashboard guidance (v2.1.198). Background agents now commit, push, and open a draft pull request when they finish code work, with progress surfaced through a Notification hook. Later builds changed the default permission mode to Manual across the CLI, VS Code, and JetBrains (v2.1.200), and made stacked slash-skill calls like /skill-a /skill-b do XYZ load every leading skill (v2.1.199).

Why it matters: Background subagents change the mental model of Claude Code from a single foreground chat to a small fleet you supervise. The Manual permission default is a safety tightening: destructive actions now wait for explicit approval instead of auto-continuing.

v2.1.197 - June 30, 2026: Claude Sonnet 5 becomes default

What changed: Claude Sonnet 5 shipped as the new default model with a 1M-token context window, alongside promotional pricing through August 31, 2026. The /model picker also gained an organization default, shown as "Org default."

Why it matters: A 1M-token window means Claude Code can hold far larger codebases and longer sessions in context without aggressive pruning, which is the single biggest lever for multi-file refactors and repo-wide reviews.

v2.1.191 - June 24, 2026: /rewind resumes past /clear

What changed: The /rewind command gained the ability to resume a conversation from before /clear was run, extending Claude Code's checkpoint system.

Why it matters: Accidental context wipes used to be unrecoverable. Rewind turns /clear into a reversible action, which pairs with the checkpointing introduced back in Claude Code 2.0.

v2.1.186 - mid-June 2026: MCP auth and a Skills tab

What changed: Anthropic added claude mcp login <name> and claude mcp logout <name> for authenticating remote MCP servers, plus a dedicated "Skills" section inside the /plugin Installed tab.

Why it matters: Named login and logout make OAuth-based MCP servers far easier to manage than juggling environment variables. See our Claude Code MCP guide for setup, and browse working servers in the MCP server directory.

v2.1.178 - June 15, 2026: agent teams simplified

What changed: The TeamCreate and TeamDelete tools were removed. Every session now has one implicit team, so you spawn teammates directly with the Agent tool's name parameter with no setup step. The same build made skills in nested .claude/skills directories load correctly and added a Tool(param:value) syntax for permission rules.

Why it matters: Team setup used to be a multi-step ceremony. Collapsing it into an implicit team removes boilerplate for multi-agent work. Our agent teams guide covers the current orchestration model in depth.

v2.1.172 - June 10, 2026: deeper subagent nesting

What changed: Subagents can now spawn their own subagents up to five levels deep, and the /plugin panel gained a search bar for marketplace plugins.

Why it matters: Five-level nesting enables genuine hierarchical delegation, where a lead agent breaks work into branches that each recruit their own helpers. Pair this with the plugin marketplace covered in our best Claude Code plugins roundup.

v2.1.161 - June 2, 2026: skills auto-load

What changed: Skills placed in .claude/skills load automatically, and skill frontmatter can declare disallowed-tools to restrict what a skill is allowed to call.

Why it matters: Auto-loading removes manual registration, and per-skill tool restrictions are a real security control. To see how skills are structured and shared, read where to find Claude Code skills.

v2.1.154 - May 28, 2026: dynamic workflows and Opus 4.8

What changed: Dynamic workflows reached general availability, letting you orchestrate tens to hundreds of agents from a single run, and Claude Opus 4.8 arrived with a fast mode. Streaming tool execution became always-on across platforms.

Why it matters: This is the release where large-scale orchestration stopped being experimental. Combined with the later agent-team simplification, it is the backbone of parallel agent work in Claude Code today.

Milestone Timeline: How Claude Code Got Here

Claude Code's major capabilities landed across 2024 and 2025 before the rapid 2.1 cadence of 2026. The table below anchors each capability to its public launch, verified against Anthropic's announcements and documentation.

For the deeper story on each capability, our everything Claude Code hub links out to focused guides on hooks, subagents, and MCP. You can also browse ready-made tools and skills that build on these features.

How to Update Claude Code

Updating Claude Code takes one command: run claude update to pull the latest release, then confirm the version with claude --version. Claude Code also self-updates on launch by default, so most users are already on a recent build.

claude update
claude --version

If an update introduces a regression, you can pin an earlier version through your package manager. When a new release deprecates a model you rely on, the /model picker now warns you and shows what the default resolves to, a change that landed in the 2.1 line.

Limitations and Tradeoffs

This page is a community-maintained summary, not Anthropic's official record, so treat the exact build dates as best-effort and confirm anything mission-critical against the official changelog. Claude Code ships multiple builds per week, which means any curated list like this one lags the newest patch by a few days.

Do not treat "newest is safest." The Manual permission default and per-skill disallowed-tools exist because more autonomy means more blast radius. Background subagents that commit and push on their own are powerful, but you should review their draft pull requests before merging rather than trusting them blindly.

Related Guides

• Everything Claude Code: The Complete 2026 Guide
• How to Install Claude Code on Mac, Windows, and Linux
• Claude Code Agent Teams: Setup and Orchestration Guide
• Best Claude Code Plugins in 2026