CrewTower Mod for Claude Code

Product Introduction

1. Definition: CrewTower is a native macOS application (macOS 15+) that functions as a centralized AI agent monitoring and interaction hub. It integrates directly with the macOS system UI, displaying a persistent, compact panel within the MacBook's notch or as a pill on other displays. Technically, it is a local client that establishes connections to supported AI coding agents (like Claude Code, Codex, Cursor) and terminal sessions to surface their status, requests, and context usage in a unified, always-accessible interface.
2. Core Value Proposition: It exists to eliminate context-switching fatigue for developers and technical users who manage multiple concurrent AI coding sessions. By aggregating real-time progress, approval requests, and contextual data from disparate apps and terminals into a single, glanceable overlay, CrewTower enables focused, uninterrupted workflow management and prompt, in-context decision-making without leaving the current window.

Main Features

1. Unified Notch-Based Session Monitor: CrewTower creates a system-level overlay that lists all active and waiting sessions from connected AI agents. Each session card displays the agent name (e.g., Claude Code, Codex), session title, elapsed time, and a visual indicator for context window usage. It uses local IPC (Inter-Process Communication) and specific agent plugins/hooks to subscribe to session events without requiring cloud syncing.
2. In-Notch Request Resolution: For supported agents (Claude Code, Codex, Cursor, Gemini CLI, Qwen Code), CrewTower intercepts specific request types—such as command execution approvals, multiple-choice questions, and plan reviews—and surfaces them directly in the notch panel. Users can approve, deny, or select answers. The app sends the response back via the agent's own local API or socket connection, not via simulated keystrokes.
3. Context-Aware Session Jump & State Recovery: The "Jump" feature utilizes macOS Accessibility APIs to programmatically locate and refocus the specific application window, terminal tab (iTerm2, Warp, Ghostty), or even tmux pane where a session originates. Furthermore, CrewTower implements a state persistence engine that caches session metadata locally, allowing it to recover and re-validate the state of recent sessions after an app or system restart.

Problems Solved

1. Pain Point: Fragmented AI Agent Management. Developers using Claude Code in the Claude app, Codex in the ChatGPT app, and CLI-based agents (opencode, Gemini CLI) in separate terminals face constant app-switching to check progress, approve actions, or provide input, leading to significant productivity loss and cognitive load.
2. Target Audience: Full-Stack and DevOps Engineers, AI-Powered Development Teams, and Technical Solo Founders. Specifically, power users who routinely operate 3+ AI coding agents simultaneously across different projects or parts of a single codebase and require oversight without breaking deep work focus.
3. Use Cases: Scenario 1: A developer has Claude Code refactoring an API module, Codex generating tests in another window, and a gemini-cli session analyzing logs in a terminal. A destructive rm -rf command request from Claude appears in the notch, is reviewed and approved without switching apps. Scenario 2: Before a meeting, a user quickly glaces at the notch to see all agent context windows (visualized as filling rings) and identifies a Claude Code session at 94% to compact it, saving token costs.

Unique Advantages

1. Strengths & Limitations (Pros & Cons):

   • Pros:
     • Zero-Context-Switch Interaction: The core innovation is enabling meaningful interaction (approvals, Q&A) from a non-intrusive, always-on-top UI element.
     • Deep, Privacy-First Integration: Operates entirely on-device; prompts, code, and transcripts never leave the Mac. Integrations work via local hooks.
     • Granular Context & Cost Awareness: Real-time visualization of per-session context window usage and aggregated provider (Anthropic, OpenAI) usage is critical for token management.
     • One-Time Purchase Model: Priced at $9.99 (early bird) as a perpetual license with updates, contrasting with the SaaS subscription norm.
   • Cons:
     • Platform & OS Lock-in: Exclusively for macOS 15+, with no support for Windows, Linux, or older macOS versions.
     • Integration Depth Variance: Full interactive support is limited to a few agents (Claude Code, Codex, Cursor, Gemini CLI, Qwen Code). Many listed integrations (GitHub Copilot, Aider, Windsurf) offer only "live activity or detection and Jump," meaning notifications without the ability to respond in-notch.
     • Requires System Permissions: Needs Accessibility access for the "Jump" feature to work reliably across terminals and apps, which may be a barrier for security-conscious users.

2. Key Alternatives & Differentiation:

   • Vs. Manual App/Terminal Switching: The obvious alternative is toggling between apps manually. CrewTower eliminates this overhead, providing aggregated status and interaction, a clear functional leap.
   • Vs. Agent-Specific Monitors (e.g., ccusage for Claude): Tools like ccusage only track usage for a single provider. CrewTower differentiates by being agent-agnostic, unifying data from Claude, OpenAI, Google, and others into one view, and adding interactive request handling.
   • Vs. Generic Menu Bar Apps (e.g., Bartender): While Bartender organizes menu bar icons, it doesn't understand or interact with the underlying AI agent sessions. CrewTower is a specialized control plane with semantic awareness of AI coding workflows, plans, and approvals.

Frequently Asked Questions (FAQ)

1. Does CrewTower work with the Claude Desktop app and ChatGPT app, or just terminals? Yes, CrewTower has dedicated integrations for the official Claude and ChatGPT desktop applications. It directly monitors Claude Code sessions and Codex tasks within these native apps, allowing you to see progress, answer requests, and jump back to the specific conversation window, not just terminal-based interactions.
2. Is my code and prompt data sent to CrewTower's servers? No. CrewTower operates on a strict local-first principle. Your prompts, code, file paths, and session transcripts are processed and displayed locally on your Mac. The app only makes outbound network calls for license validation, software updates, and to check official provider usage APIs (e.g., Anthropic, OpenAI) if you enable that feature.
3. What happens if I deny a command approval in CrewTower? Does the agent know? Yes. When you deny a command (e.g., rm -rf build/) in the CrewTower notch, it sends a structured "deny" response back to the waiting AI agent through the established local connection. The agent receives this feedback exactly as if you had typed "no" in its own interface and will typically respond by canceling the planned action or asking for an alternative.
4. Can I use CrewTower to manage AI agents on remote servers or inside Docker containers? No. CrewTower is designed to monitor and interact with processes running locally on your macOS system. It cannot directly connect to agents running on remote servers, inside virtual machines, or Docker containers, as it relies on local IPC, filesystem access, and GUI window management.
5. How does the "compact soon" warning for context windows work? CrewTower reads the available context token information provided by the agent's API (where supported). It calculates the percentage used and triggers a visual and textual warning (changing the ring color and label) when a session approaches its context limit (typically around 80-90%). This allows you to manually "compact" the session at an optimal time to retain important context before it is automatically evicted.