Claude Code Advanced Features: Subagents, Auto-Approve & Bash Mode

Master Claude Code's advanced features: subagents for parallel processing, auto-approve mode, bash execution, slash commands, and hooks. Plus how AnyCap extends capabilities beyond code.

by AnyCap

AI agent robots working in parallel on different tasks around a central hub

Claude Code is more than a terminal chatbot. Under the hood, it has a set of advanced features that transform it from a coding assistant into an autonomous development agent. Subagents, auto-approve mode, bash execution, and project-level configuration turn Claude Code into a tool that can plan, execute, and review complex multi-step tasks without you babysitting every command.

This guide covers the features that experienced Claude Code users rely on daily — and how AnyCap extends them further. New to Claude Code? Start with our installation guide and Claude Code vs Cursor comparison first.

Subagents: Parallel AI Workers

Subagents are independent Claude instances that Claude Code spawns to handle subtasks in parallel. Instead of one Claude working sequentially through a multi-file refactor, subagents divide the work.

When you ask Claude Code to "refactor the authentication module across all files," it spawns subagents — one per file or logical unit — that work simultaneously:

Main Claude session
├── Subagent 1: Refactor auth.service.ts
├── Subagent 2: Refactor auth.middleware.ts
├── Subagent 3: Refactor auth.types.ts
└── Subagent 4: Update auth tests

Each subagent gets its own context window, processes its assigned file, and returns results to the main session.

Configuration

{
  "subagents": {
    "maxConcurrent": 4,
    "maxTokensPerSubagent": 100000,
    "autoSpawn": true
  }
}

Subagents + AnyCap

Subagents handle code. AnyCap handles everything else. Combine them:

npx -y skills add anycap-ai/anycap -a claude-code

Your main Claude session orchestrates multiple subagents refactoring code while simultaneously calling AnyCap for image generation, web research, and asset storage — all in parallel. For MCP setup details, see our MCP capabilities guide.

Auto-Approve Mode: Trust But Verify (Quickly)

By default, Claude Code asks for confirmation before editing files, running shell commands, or making network requests. Auto-approve lets changes flow without confirmation pauses:

claude config set permissions.auto-accept-edits true

Granular Permissions

{
  "permissions": {
    "allow": ["read", "write"],
    "autoAccept": ["write"],
    "deny": ["shell:rm -rf *", "shell:git push --force"]
  }
}

Use auto-approve on feature branches with good test coverage. Don't use it on production infrastructure. Even with auto-approve on, /undo reverts to the last checkpoint.

Bash Mode: Full Shell Access

Claude Code can execute arbitrary shell commands — full terminal access, not a sandboxed subset:

npm install && npm run build
git diff HEAD~5 --stat
docker-compose up -d
python -m pytest tests/ --verbose

Safe Shell Configuration

{
  "permissions": {
    "deny": [
      "shell:rm -rf *",
      "shell:git push --force origin main",
      "shell:drop table"
    ]
  }
}

Bash Mode + AnyCap

AnyCap extends Claude Code's shell capabilities with managed infrastructure. Instead of manual curl for image APIs or managing cloud uploads, AnyCap handles API keys, authentication, and infrastructure:

anycap image generate --prompt "..." -o hero.png
anycap drive upload hero.png

Slash Commands: Built-in Shortcuts

Command What It Does
/init Generate CLAUDE.md from project analysis
/clear Reset conversation context
/compact Compress conversation history to free context
/cost Show token usage and session costs
/doctor Diagnose configuration issues
/mcp Manage MCP server connections
/undo Revert to last checkpoint

Custom Slash Commands

Create your own in .claude/commands/:

# .claude/commands/review.md
Review the current git diff for security vulnerabilities,
performance issues, missing error handling, and style violations.
Focus on $ARGUMENTS or the entire diff.

Now /review auth module runs your custom review scoped to the auth module.

Hooks: Automation at Lifecycle Points

{
  "hooks": {
    "postToolUse": [
      {
        "matcher": "edit_file|write_file",
        "command": "prettier --write $FILE_PATH"
      }
    ],
    "preToolUse": [
      {
        "matcher": "bash",
        "command": "echo 'Running: $TOOL_INPUT'"
      }
    ]
  }
}

Common patterns: auto-format after edits, block dangerous commands, log tool invocations, inject environment variables at session start.

Putting It All Together

A typical power-user workflow:

  1. Start a session in your project directory
  2. Subagents spin up to handle parallel refactoring tasks
  3. Auto-approve lets changes flow without confirmation pauses
  4. Bash mode runs tests, builds, and deploys
  5. Hooks auto-format every changed file
  6. AnyCap handles image generation, web research, and asset storage alongside code work
cd ~/my-saas-project
claude

> Refactor the payment module to the new Stripe API.
  Generate UI mockups for the checkout flow.
  Search for Stripe's latest pricing.
  Upload all assets to cloud storage when done.

Claude Code handles the refactoring. AnyCap handles the images, search, and storage. One session. Every capability.

You don't need subagents and auto-approve on day one. Start with basic Claude Code sessions. Add auto-approve when confirmation prompts slow you down. Enable subagents when you tackle multi-file refactors. Install AnyCap when you need capabilities beyond code. The advanced features are there when you're ready — and they transform Claude Code from a chat tool into a genuine development agent.