Claude Code Latest Features: What's New in Autumn 2025

Checkpoints: The Undo Button for AI Coding

On September 29, 2025, Anthropic introduced checkpoints, arguably the most important feature for building confidence in AI-assisted development. Checkpoints automatically save your code state before each change, enabling instant rollback when experiments go wrong.

How Checkpoints Work

Every user prompt creates a new checkpoint that persists for 30 days (configurable). When you need to rewind, press Esc twice or use the /rewind command to access three restoration options:

• Conversation only: Keep code changes, revert the chat history
• Code only: Keep conversation context, undo file modifications
• Both code and conversation: Complete rollback to a prior state

Why This Is Cool

Checkpoints solve a fundamental problem with autonomous AI coding: fear of destructive changes. Traditional version control like Git requires manual commits and discipline. Checkpoints provide instant, granular undo without ceremony. This psychological safety net enables more ambitious experimentation and delegation of complex tasks.

The system tracks only direct file edits through Claude's editing tools, not bash command modifications. This design prevents accidental checkpoint bloat from operations like rm, mv, or npm install, which should remain under explicit version control.

Real-World Use Cases

• Refactoring experiments: Try aggressive architectural changes, knowing you can instantly revert
• Bug fixes: Test multiple debugging approaches in rapid succession
• Feature iteration: Explore alternative implementations without manual Git branching

Checkpoints complement Git rather than replacing it. Use checkpoints for session-level experimentation, Git for permanent history and collaboration.

Claude Sonnet 4.5: The Engine Behind Autonomy

On September 29, 2025, Anthropic launched Claude Sonnet 4.5 as the new default model for Claude Code, calling it "the best coding model in the world." This isn't marketing hyperbole. Sonnet 4.5 demonstrates the ability to maintain focus for more than 30 hours on complex, multi-step development tasks.

Key Improvements

• Extended reasoning: Handles significantly longer task sequences without losing context
• Agent capabilities: Better at breaking down complex problems into actionable steps
• Computer use: Enhanced ability to interact with tools and interfaces
• Math and logic: Substantial gains in analytical reasoning

Pricing and Availability

Sonnet 4.5 maintains the same pricing as its predecessor: $3 input / $15 output per million tokens. Access it via the Claude API using the model ID claude-sonnet-4-5. The model is also available in GitHub Copilot as of October 13, 2025.

Impact on Claude Code

The model upgrade enables the autonomous features discussed throughout this article. Checkpoints, subagents, and extended sessions all benefit from Sonnet 4.5's improved reasoning and task persistence.

Subagents: Parallel Development Workflows

Subagents are specialised AI assistants that handle specific task types with their own context windows and tool permissions. Announced in late summer 2025 and refined through September, subagents enable true parallel development workflows.

How Subagents Work

Each subagent operates in an isolated context window with a custom system prompt and specific tool permissions. When Claude encounters work matching a subagent's expertise, it delegates the task to the specialised agent. This provides three key advantages:

• Context preservation: Main conversation doesn't get cluttered with specialised task details
• Specialised expertise: Agents can be fine-tuned with domain-specific instructions
• Parallel execution: Multiple subagents can work simultaneously on different aspects of your codebase

Creating Subagents

Use the /agents command to open an interactive interface for creating project-level or user-level agents. Subagents are stored as Markdown files with YAML frontmatter in .claude/agents/ (project) or ~/.claude/agents/ (user).

Built-in Subagent Examples

• Code reviewer: Analyses changes for quality, security, and maintainability
• Debugger: Performs root cause analysis on errors and test failures
• Data scientist: Handles SQL queries and BigQuery operations

Real-World Application

Imagine building a full-stack feature. The main agent coordinates whilst delegating backend API development to one subagent and frontend UI implementation to another. They work in parallel, each maintaining focused context on their specialised domain.

Hooks: Automated Quality Gates

Hooks are automated triggers that execute at specific lifecycle points in Claude Code sessions. Released alongside subagents, hooks enable sophisticated workflow automation without manual intervention.

Available Hook Types

• PreToolUse: Executes before Claude uses any tool, enabling approval/denial logic
• PostToolUse: Runs after successful tool execution, perfect for formatting or testing
• UserPromptSubmit: Intercepts user prompts to add context or metadata
• SessionStart: Injects environment data when sessions begin
• SessionEnd: Cleans up resources when sessions terminate

Common Hook Use Cases

• Auto-formatting: Run Prettier or PHP-CS-Fixer after code changes
• Test automation: Execute PHPUnit or Jest after modifications
• Linting enforcement: Block commits that violate ESLint or PHPStan rules
• Resource management: Control parallel execution to prevent database lock conflicts

Hook Configuration

Hooks are configured in ~/.claude/settings.json or .claude/settings.json. Each hook specifies a script path and optional metadata. The v2.0.30 release added prompt-based stop hooks, enabling even more sophisticated control patterns.

For a detailed example of using hooks to control subagent behaviour and prevent resource conflicts, see my article on Advanced Claude Code Hooks: Controlling Sub-Agent Behaviour.

Plugins: Sharing Workflows Through Marketplaces

On October 9, 2025, Anthropic launched the Claude Code plugin system in public beta. Plugins are lightweight packages that bundle slash commands, subagents, MCP servers, and hooks into shareable, installable units.

What Plugins Include

• Slash commands: Custom shortcuts for frequently-used operations
• Subagents: Purpose-built agents for specialised development tasks
• MCP servers: Connect to tools and data sources through the Model Context Protocol
• Hooks: Customise Claude Code's behaviour at key workflow points

Using the /plugin Command

Install plugins with the /plugin command. First, add a marketplace:

/plugin marketplace add user-or-org/repo-name

Then browse and install plugins:

/plugin install plugin-name@marketplace-name

Plugins work across both terminal and VS Code extension environments, providing consistent functionality regardless of interface.

Creating Plugin Marketplaces

Any Git repository can host a plugin marketplace. Create a .claude-plugin/marketplace.json file with properly formatted plugin metadata, then share the repository URL. This decentralised approach enables teams to create internal plugin marketplaces for company-specific workflows.

Why Plugins Matter

• Standardisation: Engineering leaders enforce consistency across teams
• Knowledge sharing: Open source maintainers provide best-practice workflows
• Tool integration: Connect internal tools through MCP without custom development
• Productivity patterns: Share proven debugging, testing, and deployment workflows

Agent Skills: Progressive Disclosure of Capabilities

On October 16, 2025, Anthropic introduced Agent Skills, a new pattern for making specialised abilities available to Claude models. Skills use a "progressive disclosure" design that loads information only when relevant, making the system both token-efficient and scalable.

How Skills Work

Skills are folders containing instructions, scripts, and resources. Each skill takes only a few dozen tokens in the agent's context, with full details loaded only when the user requests a task the skill can solve. Claude automatically determines which skills are relevant and loads them as needed.

Installation and Usage

Install skills via the /plugin command from the anthropics/skills marketplace, or manually by adding them to ~/.claude/skills. Skills work across Claude.ai, Claude Code, and the Claude API.

Availability

Skills are available on Pro, Max, Team, and Enterprise plans as of October 16, 2025.

Skills vs Plugins

Whilst plugins bundle multiple customisation types (commands, agents, hooks, MCP servers), skills focus specifically on specialised task capabilities. Think of skills as expertise modules and plugins as workflow packages.

VS Code Extension: Native IDE Integration

Announced on September 29, 2025, the native Claude Code VS Code extension brings AI-assisted development directly into the IDE with real-time visual feedback.

Key Features

• Inline diffs: See Claude's changes in the IDE diff viewer, not just the terminal
• Sidebar panel: Dedicated interface for Claude interactions with full chat history
• Quick launch: Press Cmd+Esc (Mac) or Ctrl+Esc (Windows/Linux) to open Claude Code
• Context awareness: Current selection and active tab automatically shared with Claude
• Plugin support: All plugins installed via /plugin work in both terminal and VS Code

Checkpoint Integration

The checkpoint system works the same way in VS Code. Press Esc twice or use /rewind to access the rewind menu, with changes displayed in the IDE's native diff viewer for clear visual feedback.

Terminal Parity

The extension maintains feature parity with the terminal interface. Subagents, hooks, skills, and plugins all function identically, ensuring consistent workflows across environments.

Status

The VS Code extension is currently in beta and available for download from the VS Code Extension Marketplace.

Claude Code on the Web and Mobile

On October 20, 2025, Anthropic expanded Claude Code beyond the terminal with a web interface and iOS app integration, making AI-assisted coding accessible without local development environments.

Web Interface Features

• GitHub integration: Connect repositories directly from claude.com/code
• Browser-based coding: No terminal or local installation required
• Sandbox security: Every task runs in an isolated environment with network and filesystem restrictions
• Git proxy service: Secure authentication ensures Claude can only access authorised repositories

Mobile Support

The iOS app enables exploratory coding on mobile devices. Whilst not intended for production development, it allows code review, bug fixes, and prototyping from anywhere.

Availability

Claude Code on the web is in research preview for Pro and Max users. Visit claude.com/code to connect your first repository.

Security Architecture

The web interface uses isolated sandbox environments for all code execution. Network and filesystem access is restricted, and Git operations go through a secure proxy that validates repository permissions. This architecture prevents malicious code execution whilst maintaining full development capabilities.

Performance Impact

According to TechCrunch, Claude Code has grown 10x in users since its broader launch in May 2025, and now accounts for more than $500 million in annualised revenue for Anthropic.

Model Context Protocol (MCP) Integration

Claude Code's support for the Model Context Protocol enables connections to hundreds of external tools and data sources through a standardized interface. MCP adoption accelerated dramatically in early 2025.

What Is MCP?

Think of MCP as "USB-C for AI." Just as USB-C provides a universal connection standard for devices, MCP provides a universal protocol for AI models to connect to different tools and services. Developed by Anthropic, MCP is an open-source standard that has seen rapid industry adoption.

Industry Adoption Timeline

• March 2025: OpenAI adopted MCP across ChatGPT
• April 2025: Google confirmed support for Gemini
• 2025: Block, Apollo, Zed, Replit, Codeium, and Sourcegraph all implemented MCP support

Available MCP Servers

Claude Code can connect to services including Stripe, Figma, Cloudinary, Canva, Sentry, Jam, Asana, and Atlassian products.

Claude Code as MCP Server

Interestingly, Claude Code can run as an MCP server itself using claude mcp serve. This exposes Claude Code's file editing and command execution tools via the MCP protocol, enabling other AI systems to use Claude Code as a tool.

Token Management

Claude Code displays warnings when MCP tool output exceeds 10,000 tokens, with a default maximum of 25,000 tokens (configurable). This prevents context window exhaustion from verbose tool responses.

Current Protocol Version

As of early November 2025, the current MCP protocol version was 2025-06-18 (a newer revision, 2025-11-25, followed later that month). The specification continues to evolve with breaking changes as it matures.

Terminal Interface 2.0

Version 2.0 of the Claude Code terminal interface, released on September 29, 2025, brings significant UX improvements for power users who prefer command-line workflows.

Key Improvements

• Enhanced status visibility: Clear indication of current agent state and active operations
• Searchable prompt history: Press Ctrl+R to search and reuse previous prompts
• Improved mode switching: Windows users can now use Shift+Tab instead of Alt+M (changed in v2.0.31)
• Better error reporting: More detailed information when operations fail

Background Tasks

The updated terminal interface supports background tasks, allowing long-running processes like development servers to remain active without blocking Claude Code's progress on other work. This enables true parallel development workflows.

Native Binary Performance

Version 2.0.33 (latest release) improved native binary installations to launch with significantly better speed, making the terminal experience even more responsive.

Configuration and Customisation Improvements

Recent releases added several quality-of-life improvements for advanced users who need fine-grained control over Claude Code's behaviour.

Security and Sandbox Controls

Version 2.0.30 introduced allowUnsandboxedCommands setting for policy-level restrictions, and disallowedTools field for custom agent definitions to explicitly block specific tools. These features enable organisations to enforce security policies whilst maintaining development flexibility.

MCP Configuration

• SSE support: Native builds now support Server-Sent Events (SSE) MCP servers (v2.0.30)
• Configuration precedence: Fixed --mcp-config flag incorrectly overriding file-based configurations (v2.0.30)
• Tool compatibility: Resolved issues with MCP tools containing incompatible output schemas (v2.0.33)

VS Code Integration Settings

Version 2.0.31 added respectGitIgnore configuration option for VS Code extension, allowing users to optionally include gitignored files in searches when needed for debugging or analysis.

Company Announcements

Version 2.0.32 introduced companyAnnouncements setting for displaying startup notifications, useful for enterprise deployments that need to communicate policy changes or updates to development teams.

Getting Started with New Features

Each feature below has a quick way to try it.

Update Claude Code

First, ensure you're running the latest version:

npm update -g @anthropic-ai/claude-code

Try Checkpoints

Start a coding session and make some changes. Then press Esc twice to open the rewind menu. Experiment with reverting code, conversation, or both to understand the workflow.

Create a Subagent

Use the /agents command to create a specialised agent for a common task in your workflow. For example, create a "test runner" agent that validates changes before commits.

Install a Plugin

Add the official marketplace and explore available plugins:

/plugin marketplace add anthropics/plugins
/plugin install your-chosen-plugin

Try the VS Code Extension

Search for "Claude Code" in the VS Code Extension Marketplace and install it. Use Cmd+Esc (Mac) or Ctrl+Esc (Windows/Linux) to launch Claude in your IDE.

Explore Claude Code on the Web

Visit claude.com/code (requires Pro or Max plan) and connect a GitHub repository. Try kicking off a coding task from your browser to experience the web interface.

Set Up an MCP Server

Browse available MCP servers and configure one relevant to your workflow. The MCP documentation provides setup instructions.

What's Next for Claude Code?

These were the trends visible from where things stood in early November 2025 - worth reading as a forecast to check against what actually shipped, not as a roadmap:

• Extended autonomous operation: With 30-hour task persistence already demonstrated, longer-running sessions with better error recovery seemed likely
• Team collaboration features: Shared plugins, subagents, and hooks looked positioned to enable team-wide consistency
• Enhanced mobile experience: The iOS app looked like a first step rather than a finished mobile workflow
• Deeper IDE integrations: the VS Code extension was still in beta, with a stable release and possible JetBrains integration plausible next
• Enterprise features: security controls and audit logging for regulated industries looked like an open gap

The pattern across that period was consistent: Claude Code moving from a coding assistant into a broader development platform spanning terminals, IDEs, web browsers, and mobile devices.

From Coding Assistant to Development Platform

Checkpoints gave a way to undo experiments, subagents split work into isolated, specialised context windows, and plugins packaged and shared both of those, plus hooks and MCP servers, as installable units. The VS Code extension and the web and mobile interfaces then moved all of that out of a single terminal window, with Claude Sonnet 4.5 as the model change underneath most of it.

Individually, several of these are conveniences, but together, over three months, they moved Claude Code from a tool you run to a platform you configure: routine tasks delegated to subagents, experiments made reversible by checkpoints, and workflows packaged as plugins rather than repeated by hand.

For developers working in PHP, TypeScript, Python, Go, or anything else, the practical starting point is checkpoints, since they cost nothing to try and remove the main reason to hold back from letting Claude Code make larger changes unsupervised.

Additional Resources

• Claude Code GitHub Repository - Source code, issues, and community discussions
• Official Changelog - Complete version history and release notes
• Claude Code Documentation - Comprehensive guides and API reference
• Anthropic News - Official announcements and feature launches
• VS Code Extension - Download and install the IDE integration
• Claude Code on the Web - Try the browser-based interface
• Official Skills Marketplace - Browse and install Agent Skills
• Model Context Protocol - Learn about MCP and available servers