Skip to main content
Pentagon needs access to local directories and gives you control over what tools each Claude Code agent can use. This guide covers both macOS-level permissions and Claude Code tool permissions.

macOS System Permissions

Pentagon accesses specific directories on your Mac. Grant these when prompted — or configure them in System Settings.

Required Directories

DirectoryPurposeAccess Type
~/.pentagon/Agent data, settings, configurationsRead + Write
~/.claude/Claude transcript files for conversation syncRead
Your project directoriesWhere agents read and write codeRead + Write
For the smoothest experience, grant Pentagon Full Disk Access:
  1. Open System Settings > Privacy & Security
  2. Click Full Disk Access in the sidebar
  3. Toggle Pentagon on (or click + and add it)
This prevents repeated permission prompts and ensures agents can access any project directory you point them at.
Pentagon shows a helpful dialog if it detects permission issues, with a button to open System Settings directly.

What Happens Without Permissions

If file access is denied:
  • Agent configurations won’t save when you quit
  • Chat history won’t persist between sessions
  • New agents may fail to create
  • Transcript syncing won’t work
  • The ~/.pentagon/ directory can’t be written to

Claude Code Tool Permissions

When spawning an agent, you choose how much autonomy Claude has over its tools.

Full Auto Mode

Sets bypassPermissions — Claude runs all tools without asking for approval.
  • Maximum agent autonomy
  • No interruptions for tool approvals
  • Best for trusted projects and experienced users

Custom Mode

Select exactly which tools the agent can use:
ToolWhat It DoesRisk Level
ReadRead files and directoriesLow
GrepSearch file contentsLow
GlobFind files by name patternLow
EditModify existing filesMedium
WriteCreate new filesMedium
BashExecute shell commandsHigh
Toggle each tool on or off in the spawn dialog. The agent can only use tools you enable.

Setting Permissions

  1. Click an empty cell or press Cmd + N to open the spawn dialog
  2. Select Full Auto or Custom under permission mode
  3. If Custom, check the tools you want to allow
  4. Click Create

Security Best Practices

Use Custom Mode for Unfamiliar Code

When working with a codebase you don’t fully trust:
  • Disable Bash to prevent arbitrary command execution
  • Enable only Read and Grep for code review tasks
  • Add Edit and Write only when you’re ready for changes

Review Bash Usage

Bash is the most powerful tool — it can run any shell command. When enabled:
  • Watch the terminal output for commands being executed
  • Consider whether the agent needs shell access at all
  • Use worktree isolation so Bash commands can’t affect your main branch

Leverage Worktree Isolation

Git worktrees act as a natural security boundary:
  • Each agent writes to its own branch
  • Changes are isolated until you explicitly merge
  • A misbehaving agent can’t corrupt your main branch
  • You can simply delete the worktree to discard all changes

Gradual Trust

If you’re new to Pentagon:
  1. Start with Custom mode — enable Read and Grep only
  2. Watch how the agent works and what it attempts
  3. Add Edit and Write when comfortable
  4. Move to Full Auto once you trust the workflow

Privacy

Pentagon keeps all data local to your Mac:
  • No cloud sync — agent data, configurations, and history stay in ~/.pentagon/
  • No telemetry — Pentagon does not collect usage data or analytics
  • No external connections — the only network traffic is between Claude Code and the Anthropic API
  • No file exfiltration — Pentagon does not access or transmit files beyond what agents need for their tasks
Your API key is stored locally and used solely for Claude Code sessions.

Next Steps