Claude Code 2026 Feature Map

1. Auto Mode Research Preview Team Plan Only

A background AI classifier (Sonnet 4.6) replaces manual permission prompts. Instead of approving every file edit or bash command, the classifier evaluates each action and blocks anything unsafe.

How It Works

• Decision order: (1) Your allow/deny rules resolve first, (2) Read-only + file edits in working dir auto-approved, (3) Everything else goes to classifier, (4) If blocked, Claude gets the reason and tries alternatives
• Stripped transcripts: Classifier sees only user messages + executable commands -- Claude's own text stripped out (prevents prompt injection via file contents)
• Reads CLAUDE.md so project-specific instructions factor into decisions

Safety Features

• Blocks: Mass deletions, data exfiltration, production deploys, force pushes to main, IAM grants, curl | bash, sending sensitive data externally
• Allows: Local file ops in working dir, installing declared dependencies, read-only HTTP, pushing to current branch
• Fallback: 3 consecutive denials or 20 total in session = auto mode pauses, resumes manual prompting
• On entering auto mode: Blanket Bash(*) allow rules get dropped. Narrow rules like Bash(npm test) carry over

How to Enable

# CLI startup
claude --enable-auto-mode
# Then cycle with Shift+Tab

# Non-interactive / scripted
claude -p "refactor the auth module" --permission-mode auto

# Settings default
{ "permissions": { "defaultMode": "auto" } }

# View default rules
claude auto-mode defaults

Availability: Team plan (research preview). Enterprise + API rolling out. Requires Sonnet 4.6 or Opus 4.6. Admins enable at claude.ai/admin-settings/claude-code.

Cost: Classifier calls count toward token usage. Extra cost mainly from shell commands and network operations.

Performance: 0.4% false-positive rate (n=10,000). 17% false-negative rate on known dangerous actions (n=52).

2. Code Review Research Preview Team/Enterprise Only

Automated PR review that dispatches a fleet of specialized agents to analyze GitHub pull requests in parallel -- looking for logic errors, security vulnerabilities, broken edge cases, and subtle regressions.

How It Works

• Multiple agents analyze the diff and surrounding code in parallel
• Verification step checks candidates against actual code to filter false positives
• Results deduplicated, ranked by severity, posted as inline comments
• Reviews complete in ~20 minutes on average

Severity Levels

• Red (Important): Bug that should be fixed before merging
• Yellow (Nit): Minor issue, worth fixing but not blocking
• Purple (Pre-existing): Bug in codebase, not introduced by this PR

Setup and Triggers

1. Admin enables at claude.ai/admin-settings/claude-code
2. Install Claude GitHub App to your org
3. Select repos and set review triggers per repo

Manual triggers: @claude review (starts review + subscribes to future pushes) or @claude review once (one-shot)

Customization: CLAUDE.md for project instructions. REVIEW.md for review-only guidance (style rules, skip patterns).

Results: Substantive review comments rose from 16% to 54% of PRs. On large PRs (>1,000 lines), 84% generated findings averaging 7.5 issues. Engineers marked less than 1% of findings incorrect.

Cost: $15-25 per review on average, billed separately through extra usage.

3. All Permission Modes Available Now

Mode	What Claude Can Do Without Asking	Best For
default	Read files	Getting started, sensitive work
acceptEdits	Read + edit files	Iterating on code
plan	Read files (proposes, doesn't execute)	Exploring, planning refactors
auto	All actions with background safety	Long-running tasks
bypassPermissions	All actions, NO checks	Isolated containers ONLY
dontAsk	Only pre-approved tools	Locked-down CI

Switch: Shift+Tab cycles modes. --permission-mode <mode> at startup.

4. Hooks System (21+ Lifecycle Events) Available Now

User-defined commands that fire at specific lifecycle points. Four handler types: command (shell), http (POST to URL), prompt (single-turn LLM), agent (multi-turn subagent).

All Hook Events (v2.1.85)

Event	When
SessionStart	Session begins or resumes
UserPromptSubmit	Prompt submitted, before processing
PreToolUse	Before a tool executes (can block)
PermissionRequest	Permission dialog appears
PostToolUse	After a tool succeeds
PostToolUseFailure	After a tool fails
Notification	Claude sends a notification
SubagentStart/Stop	Subagent lifecycle
TaskCreated/Completed	Task lifecycle
Stop	Claude finishes responding
StopFailure	Turn ends due to API error
TeammateIdle	Agent team member about to idle
InstructionsLoaded	CLAUDE.md or rules loaded
ConfigChange	Config file changes
CwdChanged/FileChanged	Reactive environment
WorktreeCreate/Remove	Worktree lifecycle
PreCompact/PostCompact	Context compaction
Elicitation/ElicitationResult	MCP input dialogs
SessionEnd	Session terminates

Example: Auto-Format on Edit

{
  "hooks": {
    "PostToolUse": [{
      "matcher": "Edit|Write",
      "hooks": [{
        "type": "command",
        "command": "jq -r '.tool_input.file_path' | xargs npx prettier --write"
      }]
    }]
  }
}

v2.1.85: Conditional if field using permission rule syntax to filter execution.

PreToolUse decisions: "allow" (skip prompt), "deny" (cancel with reason), "ask" (show prompt).

5. Voice Mode Available

/voice command activates push-to-talk. Hold spacebar to speak, release to send. 20 languages supported. Rebind key via voice:pushToTalk in keybindings.json. Requires SoX on Linux (apt install sox).

6. /loop + Scheduled Tasks Available

/loop (In-Session Recurring)

/loop 5m check the deploy -- runs every 5 minutes within session. Auto-expires after 3 days. Up to 50 concurrent.

Cloud Scheduled Tasks

Recurring tasks on Anthropic infrastructure -- even when computer is off. Create via /schedule or web at claude.ai/code/scheduled. Minimum 1 hour interval. Fresh repo clone each run. Supports MCP connectors (Slack, Linear, Google Drive).

Desktop Scheduled Tasks

Run on your machine with local file access. Minimum 1 minute interval.

	Cloud	Desktop	/loop
Runs on	Anthropic cloud	Your machine	Your machine
Machine required	No	Yes	Yes
Persistent	Yes	Yes	No (session-scoped)
Min interval	1 hour	1 minute	1 minute
Local files	No (fresh clone)	Yes	Yes

7. Agent Teams Experimental Available on Pro+

Coordinate multiple Claude Code instances working together. One lead session spawns teammates with shared task lists and direct messaging.

Enable

CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1

Architecture

• Team lead: Creates team, spawns teammates, coordinates
• Teammates: Separate instances, own context windows
• Task list: Shared with dependency tracking
• Mailbox: Inter-agent messaging
• Display: In-process (default, Shift+Down to cycle) or split panes (tmux/iTerm2)

vs Subagents

	Subagents	Agent Teams
Communication	Report back to main only	Message each other
Coordination	Main manages all	Shared task list, self-claiming
Token cost	Lower	Higher (separate instances)

8. Remote Control, Dispatch, Teleport Available on Pro+

• Remote Control: /remote-control bridges local terminal to claude.ai/code + mobile apps. Local environment stays available. Provides URL + QR code.
• Dispatch (March 17, 2026): Send tasks from phone to computer. Claude uses desktop apps to complete tasks.
• Teleport: /teleport moves session between web/iOS and local terminal with full history preserved.

9. Plugins and Marketplace Available

Extensions adding skills, agents, hooks, and MCP servers. Official marketplace: anthropics/claude-plugins-official -- 101+ plugins. 43+ community marketplaces, 834+ plugins total.

Install: /plugin marketplace add or /plugin install

Components: Skills (slash commands), Hooks (auto-fire on events), MCP servers (external connections).

10. OS-Level Sandboxing Available

Filesystem + network isolation for bash commands. Reduces permission prompts by 84%.

• Filesystem: Read everywhere (except denied), write only to CWD by default
• Network: Only approved servers
• macOS: Seatbelt framework. Linux: bubblewrap (bwrap). Windows: planned.
• Settings: sandbox.failIfUnavailable, sandbox.filesystem.allowWrite, sandbox.filesystem.denyRead

11. Effort Levels + Fast Mode + 1M Context Available

• Effort: Low / Medium / High. /effort or --effort flag. "ultrathink" = High for one turn only.
• Fast mode: /fast toggles. Same Opus 4.6, 2.5x faster output.
• 1M context window: Available on Max/Team/Enterprise. Disable: CLAUDE_CODE_DISABLE_1M_CONTEXT=1
• Max output: Opus 4.6 defaults to 64k tokens, upper bound 128k.
• Model picker: /model to switch. Left/right arrows adjust effort slider.

12. CI/CD Integrations Available

• GitHub Actions: anthropics/claude-code-action@v1 -- @claude mentions in PRs/issues trigger automated work
• Security Review Action: anthropics/claude-code-security-review -- automated OWASP scanning
• GitLab CI/CD: Community-maintained MR integration
• Slack: @Claude in channels spins up session, posts PR links back

13. Claude Agent SDK Available

Claude Code runtime as a library for building custom agents.

• Python: pip install claude-agent-sdk-python
• TypeScript: npm install @anthropic-ai/claude-agent-sdk
• Full toolset (Read, Write, Edit, Bash, Glob, Grep, WebSearch, WebFetch)
• Custom tools as Python functions (in-process MCP servers)
• Hooks, structured outputs, agent loop built-in
• betas option for 1M context window

14. Channels Research Preview

MCP servers that push messages INTO a Claude Code session from Telegram, Discord, iMessage, or custom webhooks. Enable: --channels flag at startup. When active, AskUserQuestion and plan-mode tools are disabled. Permission relay forwards approval prompts to phone.

15. Worktrees Available

--worktree (-w) creates isolated repo copy under .claude/worktrees/ with own branch and git index. Subagents support isolation: "worktree". worktree.sparsePaths for large monorepos. ExitWorktree tool to leave. Resuming a session auto-switches back to its worktree.

16. Desktop App Available

Standalone application: macOS (Intel + Apple Silicon), Windows (x64). Visual diff review, multiple sessions side by side, scheduled tasks, /desktop handoff from terminal.

Computer Use: Claude can screenshot, point, click, and navigate your screen when no tool access exists for a task.

17-25. Additional Features

Feature	Status	Details
Auto-Memory	Available	Auto-saves context to MEMORY.md. /memory to manage. 25KB/200 line limit.
MCP Enhancements	Available	Elicitation, OAuth (RFC 9728), tool lazy loading (95% context savings), 2KB desc cap, headers helpers
Checkpoints	Available	Auto save points before each change. Esc-twice or /rewind.
PowerShell	Preview	Windows native PowerShell tool (opt-in, v2.1.84)
Deep Links	Available	Queries up to 5,000 chars, open in preferred terminal
/btw Command	Available	Side questions that don't interrupt main workflow
/branch (fork)	Available	Fork conversation to explore alternatives
Transcript Search	Available	Press / in transcript mode, n/N to navigate
Security Review	Available	/security-review for pre-commit OWASP scanning

Current Plan: Claude Max 20x ($200/mo)

Feature	Access	Notes
Claude Code CLI	YES	Full access
Opus 4.6	YES	Default model
1M Context Window	YES	Max plan perk
Fast Mode	YES	/fast
Voice Mode	YES	/voice
Agent Teams	YES	Experimental flag
Subagents	YES	Built-in
Hooks (21+ events)	YES	Full access
Worktrees	YES	--worktree
/loop + Scheduling	YES	Cloud + Desktop
Remote Control	YES	/remote-control
Dispatch	YES	Phone to Computer
Teleport	YES	/teleport
Plugins	YES	/plugin
Sandboxing	YES	bwrap on Linux
Agent SDK	YES	Python + TypeScript
GitHub Actions	YES	claude-code-action
MCP Enhancements	YES	All features
Channels	YES	Research preview
Checkpoints	YES	Esc-twice or /rewind
Auto-Memory	YES	Already using
Auto Mode	TEAM ONLY	Research preview, needs Team plan
Code Review (PR)	TEAM ONLY	Team/Enterprise only
Admin Controls	TEAM ONLY	claude.ai/admin-settings
Managed Settings	TEAM/ENT	Organization policies

Variable	Purpose
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS	Enable agent teams
CLAUDE_CODE_DISABLE_1M_CONTEXT	Disable 1M context
CLAUDE_CODE_DISABLE_CRON	Disable cron scheduling
CLAUDE_CODE_MAX_OUTPUT_TOKENS	Override max output tokens
CLAUDE_STREAM_IDLE_TIMEOUT_MS	Streaming idle watchdog (default 90s)
CLAUDE_CODE_SUBPROCESS_ENV_SCRUB	Strip credentials from subprocesses
ANTHROPIC_DEFAULT_MODEL_NAME	Customize model picker labels
ANTHROPIC_CUSTOM_MODEL_OPTION	Custom /model entries
ENABLE_CLAUDEAI_MCP_SERVERS	Toggle claude.ai MCP servers
OTEL_LOG_TOOL_DETAILS	Gate tool parameters in OpenTelemetry

Command / Flag	Description
/plan	Enter plan mode for single request
/voice	Activate voice mode
/loop	Recurring prompt at intervals
/schedule	Create cloud scheduled task
/copy	Interactive picker for code blocks
/color	Set session prompt-bar color
/rename	Rename session
/effort	Set model effort level
/context	View context usage + optimization tips
/hooks	Browse configured hooks
/branch	Fork conversation
/compact	Manual context compaction
/export	Export session
/teleport	Transfer session CLI to web
/desktop	Hand off to Desktop app
/remote-control	Bridge to claude.ai/code
/btw	Side question during response
/fast	Toggle fast mode
/model	Switch models
/status	Show status during response
/debug	Toggle debug mode
/memory	Manage auto-memory
/security-review	Pre-commit security scan
/plugin	Plugin management
--worktree / -w	Git worktree isolation
--channels	Enable channel servers
--enable-auto-mode	Make auto mode available
--permission-mode	Set starting permission mode
--print / -p	Non-interactive/headless mode
--bare	Skip hooks/LSP/plugins for scripts
--json-schema	Structured output
--resume / --continue	Resume previous session