Claude Code in 2026: Voice Mode, /loop, Scheduled Tasks, and Everything New

#Claude Code Just Got Serious

Claude Code has gone from "smart terminal assistant" to a full-blown autonomous development platform in early 2026. If you haven't kept up with the changelog, you've missed some major shifts in what this tool can do.

This guide covers everything that's landed since the start of 2026 — the features that actually change how you work, not just incremental fixes. We'll focus on practical usage: what each feature does, when to use it, and how to set it up.

Let's get into it.

#Voice Mode — Talk to Your Code

Shipped: March 3, 2026 (rolling out gradually)

Claude Code now has native voice input. Instead of typing every command, you can speak naturally and Claude executes.

#How to Enable It

# Toggle voice mode on/off
/voice

When voice mode is active, you'll see an indicator in the terminal. Hold to speak (push-to-talk), release, and Claude transcribes and executes your request.

#What It's Good For

• Hands-free refactoring: "Refactor the authentication middleware to use JWT tokens"
• Quick navigation: "Show me all the files that import the database module"
• Reviewing while multitasking: "Summarize the changes in the last three commits"
• Accessibility: For developers who find extended typing uncomfortable

#What to Know

• Currently live for ~5% of users, ramping through March 2026
• You'll see a note on the welcome screen once you have access
• Push-to-talk keybinding is rebindable (as of v2.1.71+)
• Transcription accuracy has been specifically improved for developer terms — regex, OAuth, JSON, and repo names are recognized correctly (v2.1.72)
• On macOS native binary, make sure your terminal has microphone permission — this was silently failing until v2.1.74 fixed it

#Tips

• Voice works best for high-level instructions. "Add error handling to the payment flow" is better than dictating exact code
• Combine with /plan mode: speak your plan, let Claude outline it, then approve
• If you get "No speech detected" errors, update to v2.1.72+ — this was a known bug

#/loop — Short-Term Recurring Tasks

Shipped: v2.1.71 (early March 2026)

The /loop command creates cron-like recurring tasks inside your active session. Think of it as a lightweight automation layer for things you need to watch for a few hours or days.

#How to Use It

# Start a loop
/loop
# Claude will ask what you want to repeat and how often
# Example: "Check if the CI pipeline passes every 15 minutes"
# Example: "Monitor my inbox for emails from the client every hour"

Under the hood, /loop creates a cron job tied to your session. It runs your instruction at the interval you specify.

#Key Constraints

• Maximum lifetime: 3 days — loops auto-expire after this
• Session-bound: If you close the session, the loop is gone. Missed runs don't catch up
• No cross-session persistence: Each loop lives and dies with its session

#When to Use /loop

#When NOT to Use /loop

• Daily standups or recurring reports → Use Scheduled Tasks instead
• Anything that needs to survive a reboot → Not /loop
• Long-term monitoring (weeks/months) → Use proper cron or OpenClaw heartbeats

#Managing Loops

# List active loops
cron list
# Delete a specific loop
cron delete <id>
# Kill all cron jobs in session
# Set env var: CLAUDE_CODE_DISABLE_CRON=true

#Scheduled Tasks — Persistent Automation

Scheduled Tasks are the durable counterpart to /loop. They create repeatable runs that persist across sessions, start a fresh instance each time, and process your project files and skills.

#How They Work

1. Set up through the Claude Code Desktop app (not terminal or VS Code — yet)
2. Choose your schedule: hourly, daily, weekly, or custom cron expression
3. Pick a working directory and execution preferences
4. The task runs at the specified time, reads your project, applies skills, and completes

#Key Differences from /loop

#Tradeoffs

• Your computer must be on and the Desktop app open for tasks to run on schedule
• Missed runs do get queued and execute when you reopen — much more resilient than /loop
• Each run is isolated, so there's no shared state between executions

#Best Uses

• Daily code review digest: Summarize new PRs every morning
• Weekly dependency audit: Check for outdated or vulnerable packages
• Hourly log monitor: Watch for errors in production logs
• Daily test suite: Run tests and report failures before standup

#Auto-Memory — Claude Remembers

Shipped: v2.1.59

Claude Code now automatically saves useful context from your sessions to memory. No more losing important details when a session ends.

#How It Works

• Claude identifies useful context during conversations (project patterns, your preferences, architecture decisions) and saves them to auto-memory
• Manage what's been saved with /memory
• As of v2.1.74, you can configure a custom directory for auto-memory storage with the autoMemoryDirectory setting

#Practical Impact

After a few sessions in a project, Claude will remember:

• Your code style preferences
• Architecture patterns in the codebase
• Testing conventions
• Common file locations and project structure
• Previous decisions and their rationale

This compounds over time. Session 10 is meaningfully better than session 1.

#Agent Teams & Worktrees — Parallel Work

Claude Code now supports multi-agent workflows with isolated git worktrees, meaning Claude can work on multiple things in your repo simultaneously without conflicts.

#Worktrees

# Start a session in an isolated worktree
claude --worktree
# or
claude -w

Each worktree gets its own branch and working directory. Perfect for:

• Working on a feature while Claude handles a separate bug fix
• Running parallel experiments without branch conflicts
• Subagents that need to make changes without affecting your main branch

#Agent Teams

• Define agents with background: true in agent definitions
• Agents can run in isolated worktrees with isolation: worktree
• Kill background agents with Ctrl+F (two-press confirmation)
• Per-invocation model overrides restored on the Agent tool (v2.1.72) — run different agents on different models

#New in v2.1.72+

• ExitWorktree tool to cleanly leave a worktree session
• WorktreeCreate and WorktreeRemove hook events for automation
• Background bash processes spawned by subagents are now cleaned up on exit (v2.1.73)

#Skills 2.0 — Measurable, Testable Skills

Skills have been upgraded with evaluation capabilities. Instead of guessing whether a skill is helping, you can now measure it.

#What's New

• Skill scoring: Skills can include test cases and evaluation criteria
• Skill Creator: Built-in tool to design, structure, and test skills
• Hooks integration: Skills can fire hooks on specific events
• Allowed tools: Skills can declare which tools they need access to

#Why It Matters

Before Skills 2.0, you had to trust that a skill was working. Now you can:

1. Define expected inputs and outputs
2. Run evaluations to verify behavior
3. Score skill performance over time
4. Share skills with confidence that they work as advertised

#/context — Know What's Eating Your Tokens

New in v2.1.74

The /context command now gives actionable suggestions about your session's resource usage:

• Identifies context-heavy tools consuming disproportionate tokens
• Warns about memory bloat from large auto-memory or skill files
• Shows capacity warnings with specific optimization tips

This is essential for long sessions where you're hitting context limits. Instead of guessing why Claude is forgetting things, /context tells you exactly where your tokens are going.

#Effort Levels — Simplified

Changed in v2.1.72

Effort levels have been simplified from four options to three:

# Set effort level
/effort low
/effort medium
/effort high
# Reset to automatic
/effort auto

The max level has been removed. You can now change effort while Claude is responding (just like /model).

#Other Notable Changes (March 2026)

Quality of life improvements you'll actually notice:

• Opus 4.6 is now the default on Bedrock, Vertex, and Microsoft Foundry (was Opus 4.1)
• Full 1M context window included in Opus 4.6 fast mode (v2.1.50)
• /plan now accepts a description — /plan fix the auth bug enters plan mode and starts immediately
• /copy now has a w key to write selections directly to a file (huge for SSH)
• CLAUDE.md HTML comments are now hidden from Claude when auto-injected, keeping your context cleaner
• Up arrow after interrupting restores your prompt and rewinds the conversation
• RTL language support — Hebrew, Arabic, and other RTL text now renders correctly
• Plugin marketplace improvements — submodule syncing, custom npm registries, version pinning
• modelOverrides setting — map model picker entries to custom provider model IDs (Bedrock ARNs, etc.)
• Expanded bash auto-allowlist — lsof, pgrep, tput, ss, fd, fdfind no longer trigger permission prompts

#For OpenClaw Users: How This Fits Together

If you're running OpenClaw, Claude Code's new features complement your agent setup:

Use /loop for development sprints: Your OpenClaw agent handles daily operations (email, calendar, monitoring). Use Claude Code's /loop for code-specific watching during active development — CI checks, test runs, deployment monitoring.

Use Scheduled Tasks for code maintenance: Set up daily dependency checks, weekly security audits, or hourly log monitoring through Claude Code's scheduled tasks. Reserve OpenClaw's cron for non-code tasks.

Voice mode + OpenClaw messaging: Talk to Claude Code while your OpenClaw agent sends you updates via Telegram/Discord. Two agents, two channels, zero conflicts.

Agent Teams for complex projects: Spawn Claude Code worktree agents for parallel feature work while your main OpenClaw agent handles project management, documentation, and stakeholder communication.

#Quick Start Checklist

Ready to try everything? Here's your setup order:

1. Update Claude Code — npm update -g @anthropic-ai/claude-code (make sure you're on v2.1.74+)
2. Try voice mode — Type /voice and check if you have access
3. Set up a /loop — Pick something you're already checking manually and automate it
4. Review /context — See where your tokens are going
5. Configure auto-memory — Run /memory to see what Claude has learned about your project
6. Explore agent teams — Try claude -w for your next parallel task

Claude Code in 2026 isn't just a coding assistant anymore. It's a development platform that remembers, automates, and scales. The tools are there — the question is how far you push them.

For more OpenClaw guides, check out our full guide library. Need help configuring skills for Claude Code? Browse our skill library for ready-to-use templates.