CrewTower

Product Introduction

1. Definition: CrewTower is a native macOS desktop application (Swift/SwiftUI) that functions as a centralized monitoring and control panel for AI-powered coding agents. It operates as a persistent, system-level overlay that integrates with the MacBook's physical notch or displays as a floating pill on external monitors.
2. Core Value Proposition: It exists to eliminate context-switching and silent stalls when managing multiple concurrent AI coding sessions (like Claude Code, Cursor, or Codex). Its core value is providing real-time, contextual oversight and approval workflows directly from the user's primary line of sight (the menu bar/notch), without requiring terminal interaction.

Main Features

1. Notch-Based Live Control Panel: The application renders a compact, interactive status bar within or adjacent to the MacBook's notch. It displays all active AI agent sessions, showing their name (e.g., project directory), current activity (editing, running tests), time elapsed, and a visual indicator of remaining context window capacity (with color-coded warnings at 70% and 90% usage). Sessions awaiting user input automatically rise to the top of the list.
2. Socket-Based Permission & Query Resolution: Instead of simulating keystrokes, CrewTower installs lightweight observation hooks into supported AI agent configs (e.g., ~/.cursor/hooks.json). When an agent requires approval or a decision, its hook blocks on a local Unix domain socket. CrewTower sends the user's decision (Approve, Deny, or a selected multiple-choice answer) back through this socket, resolving the prompt directly within the agent's process. This ensures security and prevents input injection into unintended windows.
3. Precise Terminal Session Jumping: Using macOS Accessibility APIs (with user permission), the application can map a displayed session to its exact originating terminal tab (in Terminal, iTerm2, Warp, etc.) or tmux pane. Clicking on a session brings that specific tab to the foreground, enabling quick context recovery. The matching is based on the unique title the AI agent assigns to its own terminal session.
4. Intelligent Notification & Privacy System: The system uses synthesized audio cues and visual "nudges" to alert users to finished sessions or long-pending requests, with escalating reminders at 1, 5, and 15-minute intervals. A dedicated "Privacy Mode" disables all visual expansions and sounds (ideal for screen sharing) while the backend monitoring continues to function.
5. Local-First, Privacy-Centric Architecture: All data processing occurs on-device. The app reads from local agent transcript files and communicates only over local user-restricted sockets. The only external network calls are for license validation, update checks, and (if enabled) fetching Claude API quota usage from locally-stored Anthropic credentials. No code, prompts, or project data is transmitted.

Problems Solved

1. Pain Point: Context Switching Overhead and "Silent Stalls." Developers lose focus and productivity when they must manually alt-tab to multiple terminal windows to check which AI agent is stuck, what it is asking for, or if it has finished. An agent blocked on a permission request produces no alert, leading to wasted minutes or hours of idle time.
2. Target Audience: Professional software developers and engineers who routinely use multiple AI coding assistants (Claude Code, Cursor with Composer, GitHub Copilot Chat, etc.) concurrently on different tasks or projects. This includes full-stack developers, DevOps engineers, and technical leads managing complex codebases where automated agent work is parallelized.
3. Use Cases: Parallel AI-Driven Development Sprints: Running one agent on API refactoring, another on frontend component generation, and a third on writing tests, all simultaneously. High-Stakes Operations: Safely approving potentially destructive shell commands (e.g., rm -rf, database migrations) with full context before execution. Resource Management: Monitoring context token usage across several long-running agent sessions to prevent unexpected truncation mid-task.

Unique Advantages

1. Differentiation: Unlike simple terminal multiplexers or standalone agent UIs, CrewTower is agent-agnostic and system-integrated. It does not replace the terminal or the AI tool; it becomes a unified layer of observability and control across all of them. Compared to checking individual terminal tabs, it provides a consolidated, prioritized, and always-visible dashboard.
2. Key Innovation: The non-invasive hook-and-socket communication model is its core technical innovation. By leveraging the agents' own hook systems, it intercepts requests at the process level without scraping the terminal or simulating UI events. This makes interactions secure, reliable, and transparent to the underlying agent workflow. The notch-as-UI paradigm is a secondary innovation in spatial interface design for developer tools.

Frequently Asked Questions (FAQ)

1. Is CrewTower compatible with Macs without a notch? Yes, CrewTower is fully compatible with all supported Macs. On MacBooks without a notch or when using an external display, it automatically renders as a floating, centered pill window at the top of the screen, maintaining identical functionality and geometry.
2. How does CrewTower handle my data and privacy? CrewTower operates on a strict local-first principle. Your source code, AI prompts, and session transcripts never leave your computer. It reads from local files created by the AI agents and communicates via local Unix sockets. The only external communications are for app updates, license checks, and optional Claude API quota checks using your locally stored credentials.
3. What happens if I quit CrewTower while AI agents are running? The application is designed to fail gracefully. For agents like Claude Code and Cursor, the installed hooks will simply return "no decision," causing the agent to fall back to its standard behavior (typically showing a prompt in the terminal). For agents like Gemini CLI that interpret no response as approval, the hook is designed to return an explicit denial to prevent unauthorized actions.
4. Can I use CrewTower with any terminal or AI agent? It has built-in support for major terminals (Terminal, iTerm2, Warp, etc.) and AI agents (Claude Code, Cursor, Codex, etc.). For unsupported terminals, you can define a custom matching rule with a one-line configuration. Unrecognized CLI agents will still be tracked and allow you to jump to their terminal tab.
5. Does CrewTower require an internet connection to function? No, its core monitoring, approval, and control functions work entirely offline. An internet connection is only required for the initial license activation, periodic license validation (approximately once daily), checking for application updates, and the optional Claude quota display feature.