Beta Features and Early Access

Claude Code continues to evolve rapidly with experimental features and beta programs. This guide explains how to access, test, and provide feedback on cutting-edge capabilities before they reach general availability.

Understanding Feature Stages

Alpha

Internal or limited testing with significant changes expected. May have bugs or incomplete functionality.

Beta

Feature-complete but gathering feedback. Available to specific user groups or plans.

Preview

Near-final implementation. Broadly available but may have minor adjustments.

GA

Generally Available. Stable, supported, and recommended for production use.

Current Beta Features

Agent View Research Preview New in v2.1.139

The full-terminal dashboard for every background Claude Code session on your machine — dispatch, peek, attach, detach, all in one screen. Debuted on May 11, 2026.

# Open the dashboard
claude agents

# Dispatch a background session from the shell
claude --bg "investigate the flaky SettingsChangeDetector test"

# Or background the current session from inside it
/bg

Available on: Pro, Max, Team, Enterprise, and Claude API plans.

Highlights:

• Per-user supervisor process keeps sessions alive when no terminal is attached
• Edits auto-isolated into .claude/worktrees/ (git repos only)
• Row-level pull request status with hyperlinked dots
• Filter syntax (a:<name>, s:<state>, #<PR>) for surfacing what you care about

Limitations: sessions run locally and stop when the machine shuts down (they recover across ordinary sleep/wake), rate limits sum across parallel agents, and deleting a session also removes a Claude-created worktree unless the shell cleanup path preserves uncommitted changes.

To turn off: set disableAgentView: true in settings or CLAUDE_CODE_DISABLE_AGENT_VIEW=1.

Tip

Copy-paste prompt to dispatch a self-contained background investigation you can check on later from claude agents:

claude --bg "Reproduce the intermittent 500 on POST /api/checkout. Add a failing integration test that triggers it, find the root cause, fix it, and run the full test suite. Stop only when the new test plus the existing suite are green. Leave a one-paragraph summary of the root cause."

Full walkthrough including keyboard shortcuts, dispatch prefixes, and the supervisor architecture lives at Agent View: claude agents.

Goal Workflows with /goal New in v2.1.139

Not strictly a research preview — /goal shipped GA in v2.1.139 — but listed here because it’s the autonomous-work counterpart to agent view and the two are designed to pair.

/goal all tests in tests/auth pass and npm run lint exits 0

Claude works turn after turn until a small fast model (default Haiku) confirms the condition. Built on a session-scoped prompt-based Stop hook; requires a trusted workspace; respects disableAllHooks and allowManagedHooksOnly.

Pair with claude --bg to dispatch goal-driven background sessions you can come back to.

Tip

Copy-paste prompt for a verifiable /goal condition - phrase it as a checkable end state, not a task, so the watcher model can confirm completion:

/goal npm run typecheck exits 0, npm run lint exits 0, and every test in tests/billing passes. Do not edit test files to make them pass.

Full guide: Goal workflows.

Agent Teams Research Preview

Multi-agent collaboration where Claude Code spawns and coordinates teammate agents:

# Enable agent teams (research preview)
export CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1

Features:

• Spawn multiple agent teammates that work in parallel
• TeammateIdle and TaskCompleted hooks for coordination (v2.1.33)
• Agents can have persistent memory with user, project, or local scope
• Restrict subagents with Agent(worker, researcher) in a main agent’s tools frontmatter, or deny individual types with permissions.deny: ["Agent(name)"]

Caution

Agent teams is a research preview and is token-intensive. Requires setting the environment variable above to enable. Available since v2.1.32.

Fast Mode New in v2.1.36

Fast mode uses the same Claude Opus model with faster output — no model switch. On Opus 4.8 it runs at 2× the standard rate for 2.5× the speed.

# Toggle fast mode inside a session
/fast

Note

Fast mode currently supports Opus 4.8 and Opus 4.7. Opus 4.7 fast mode was deprecated on June 25, 2026 and is scheduled for removal on July 24; migrate to Opus 4.8. It is not available on Sonnet, Haiku, Fable, or Opus 4.6.

GitHub Actions Integration Beta

Claude Code GitHub Actions enables AI-powered automation in your GitHub workflow:

.github/workflows/claude-review.yml

name: Claude Code Review
on: [pull_request, issue_comment]

jobs:
  claude-action:
    runs-on: ubuntu-latest
    steps:
      - uses: anthropics/claude-code-action@v1
        with:
          github_token: ${{ secrets.GITHUB_TOKEN }}
          anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}

Features:

• Respond to @claude mentions in issues and PRs
• Automatic code review and implementation
• Follows repository’s CLAUDE.md guidelines
• Creates branches, commits, and PRs automatically

Note

GitHub Actions integration is built on the Claude Code SDK, allowing custom automation workflows beyond standard features.

Native Binary Installation Default

Since v2.1.113 (April 17, 2026) the native binary replaced the bundled JavaScript build as the primary, generally-available install method. It’s the path Anthropic tests and supports, so new installs default to it:

Install

# From existing Claude Code installation
claude install

# Fresh install via script
curl -fsSL claude.ai/install.sh | bash

Platforms

• macOS: Universal binary (Intel + Apple Silicon)
• Linux: x64 and ARM64 support
• Windows: Native PowerShell and CMD installers plus WinGet; WSL remains optional for Linux toolchains

Benefits

• Faster startup times
• Smaller memory footprint
• Better system integration
• No Node.js dependency

Note

On Windows, Claude Code runs natively from PowerShell or CMD. Git for Windows is optional (without it Claude uses the PowerShell tool); WSL is required only when you want a Linux environment or sandboxed command execution. The npm package remains a supported advanced install path and requires Node.js 22+ as of v2.1.198.

Figma Dev Mode MCP Server Beta

Design-to-code integration through Model Context Protocol:

# Enable in Figma Desktop App
# Preferences -> Dev Mode -> Enable MCP Server

# Add the local desktop server to Claude Code (HTTP, /mcp endpoint)
claude mcp add --transport http figma-dev-mode http://127.0.0.1:3845/mcp

# Or use Figma's hosted remote server (no desktop app required)
claude mcp add --transport http figma-remote-mcp https://mcp.figma.com/mcp

Capabilities:

• Generate code from Figma designs
• Extract design tokens and variables
• Access Code Connect mappings
• Screenshot design regions

Caution

Requires Figma Dev Mode seat on Professional, Organization, or Enterprise plans.

Experimental Flags and Features

Environment Variables

CLAUDE_CODE_SHELL

Override automatic shell detection (v2.0.65)

export CLAUDE_CODE_SHELL=/bin/zsh

CLAUDE_CODE_PROXY_RESOLVES_HOSTS

Enable proxy DNS resolution (opt-in since v2.0.55)

export CLAUDE_CODE_PROXY_RESOLVES_HOSTS=true

CLAUDE_BASH_MAINTAIN_PROJECT_WORKING_DIR

Lock Bash tool to project root directory

export CLAUDE_BASH_MAINTAIN_PROJECT_WORKING_DIR=1

MCP_TIMEOUT

Adjust MCP server startup timeout (milliseconds)

export MCP_TIMEOUT=30000

Note

Thinking mode configuration has moved to /config as of v2.0.67. The DISABLE_INTERLEAVED_THINKING environment variable is deprecated.

Model Context Protocol (MCP) Servers

The MCP ecosystem includes many production-ready servers:

Development Tools

• Playwright/Puppeteer: Browser automation and testing
• Git: Version control operations
• GitHub/GitLab: Repository management
• PostgreSQL/MySQL: Database operations

Documentation & Knowledge

• Context7: Up-to-date library documentation
• Wikipedia: Reference information
• Google Search: Web search capabilities

Productivity

• Linear/Jira: Task management
• Slack: Team communication
• Google Drive: Document access
• Notion: Knowledge base integration

Tip

MCP servers are not experimental - they’re a core feature of Claude Code. New servers are regularly added by the community.

Accessing Features

Subscription Access

Claude Code currently supports:

1. Claude Pro: Sonnet 5 account default with included usage subject to plan limits
2. Claude Max: Higher included limits for sustained agentic sessions, with Opus 4.8 as the account default
3. Team Standard / Premium: Centralized billing and administration; the seat type determines the default model and limits
4. Enterprise subscription / pay-as-you-go: SSO, managed policies, model controls, and the billing-specific default model
5. Anthropic API and cloud providers: Usage-based billing through Anthropic, Amazon Bedrock, Google Cloud’s Agent Platform, Microsoft Foundry, or Claude Platform on AWS

Note

Claude Sonnet 5 became the account default for Pro, Team Standard, and Enterprise subscription seats on June 30, 2026. Opus 4.8 is the default on Max, Team Premium, Enterprise pay-as-you-go, Anthropic API, Amazon Bedrock, Google Cloud’s Agent Platform, and Claude Platform on AWS; Microsoft Foundry defaults to Sonnet 4.5. Claude Fable 5 is never the automatic account default; it returned globally on July 1 and, after temporary inclusion through July 7, uses usage credits on eligible paid plans. It is unavailable under zero data retention. See model comparison for details.

Checking Your Version

# View current version and features
claude --version

# Check system information
claude doctor

# View available CLI flags
claude --help

# Check MCP server status (terminal command)
claude mcp list

Recent Features (2025-2026)

Based on recent releases, these features have graduated from beta to GA:

Claude Fable 5

Released in v2.1.170 (June 9, 2026) - New tier above Opus 4.8; switch with /model fable. Restored globally July 1; after temporary inclusion through July 7, it uses usage credits.

Claude Sonnet 5

Released June 30, 2026; account default on Pro, Team Standard, and Enterprise subscription seats, with native 1M context

Claude Code for Desktop

Released in v2.0.51 - Standalone desktop application (Download)

Background Agents

Released in v2.0.60 - Run agents in background while you work

Named Sessions

Released in v2.0.64 - Use /rename and /resume <name> for session management

Rules Directory

Released in v2.0.64 - New .claude/rules/ directory for memory organization

Prompt Suggestions

Released in v2.0.67 - Tab to accept AI-suggested prompts

Model Switching

Released in v2.0.65 - Switch models with alt+p / option+p while typing

Claude Opus 4.6

Released in v2.1.32 - frontier model at the time, since superseded by Opus 4.7 (v2.1.111) and the current Opus 4.8 (v2.1.154)

Automatic Memories

Released in v2.1.32 - Claude records and recalls memories automatically

Hooks System

Released in v1.0.38 - User-defined automation points

Community-Driven Innovation

Contributing Ideas

1. GitHub Issues: Report bugs and request features at the official repository
2. Community Forums: Share use cases and workflows
3. Direct Feedback: Use /feedback command in Claude Code
4. Pull Requests: Contribute code directly to open-source components

Testing Features Safely

Isolated Testing

# Create test environment
mkdir claude-test && cd claude-test
git init

# Test with a fresh project directory
claude

# Use temporary settings directory
export CLAUDE_CONFIG_DIR=~/.claude-test

Feature Flags

// In ~/.claude/settings.json (user scope)
{
  "env": {
    "DISABLE_AUTOUPDATER": "1"
  }
}

# Or use environment variable for debug logging
export ANTHROPIC_LOG=debug

Providing Feedback

Effective Bug Reports

1. Environment Details

   claude doctor > diagnostic.txt

2. Reproduction Steps

   • Minimal test case
   • Expected vs actual behavior
   • Error messages and logs

3. Submit via Appropriate Channel

   • GitHub Issues for bugs
   • /feedback for suggestions
   • Community forums for discussion

Best Practices for Beta Testing

Version Control

Always commit before testing new features

Incremental Testing

Start with non-critical projects

Document Findings

Keep notes on behavior changes

Share Knowledge

Help others with your discoveries

Troubleshooting Beta Features

Common Issues

MCP Server Connection Problems:

# Debug MCP connections
claude --debug "mcp"

# List configured MCP servers
claude mcp list

# To fix a broken server, remove and re-add it.
# Options (--transport, --env, --scope) go BEFORE the name;
# the `--` separates the name from the command and its args.
claude mcp remove <server-name>
claude mcp add <server-name> -- <command> [args...]

Permission Issues:

# View and manage permissions interactively
/permissions
# Use the interactive prompt to allow or deny specific tools

Model Access:

# Check available models (inside a REPL session)
/model

# Upgrade plan if needed (inside a REPL session)
/upgrade

Note

Remember: Features evolve rapidly. What’s beta today may be GA tomorrow. Stay updated with the latest release notes and official announcements.