Cubicle

Product Introduction

1. Definition: Cubicle is a real-time, browser-based visualization dashboard and monitoring tool for AI agent workflows. Technically, it is a lightweight, dependency-free Node.js server that serves a pixel-art simulation interface, consuming standardized JSON feeds from AI agent platforms.
2. Core Value Proposition: It exists to provide developers and teams with an intuitive, at-a-glance operational intelligence layer for their AI agent systems, transforming abstract logs and status codes into a live, animated office environment where agent activity, bottlenecks, and performance are immediately visible.

Main Features

1. Live Pixel-Art Office Simulation: The core interface is a real-time animated office where each AI agent is represented by a pixel-art character. Agents walk to desks, type when working, raise a hand when awaiting human input, and return to a lounge when idle. This visualization works by polling a configurable data source (like a Paperclip instance API or a local JSON file) and mapping agent statuses (running, waiting, error, idle) to specific character animations and screen colors.
2. Multi-Source Agent Integration: Cubicle is platform-agnostic. It includes built-in adapters for Paperclip, Claude Code, Codex CLI, and Gemini CLI. For any other system, it can ingest a simple, documented JSON feed format (FEED.md), allowing it to visualize agents from custom scripts, multiple platforms simultaneously (e.g., --source paperclip,claude-code), or even replay recorded sessions.
3. Operational Dashboard & Alerting: Beyond visualization, it provides a KPI board showing task progress, subtask completion, and estimated finish times. It features optional browser desktop notifications and sound alerts for agents requiring attention. An integrated Telegram bot can forward agent questions to a user's phone and post the user's reply back as a comment, creating a lightweight mobile interaction layer.
4. Customizable Themes and Branding: Users are not locked into one visual style. Cubicle supports multiple high-definition theme packs (HQ, plaza, warehouse, factory, space base, military ops) alongside the default pixel theme. It allows deep customization: agent appearances can be set using a built-in wizard to crop faces from photos, and a company logo can be displayed on the office wall.
5. Zero-Dependency, Read-Only Architecture: The entire tool consists of a single Node.js server file and static frontend assets. It has zero external npm dependencies, making deployment trivial via npx. It is designed as a read-only observer; it only makes GET requests to a defined allowlist and does not execute commands or alter the agent systems it monitors, ensuring operational safety.

Problems Solved

1. Pain Point: Lack of intuitive visibility into distributed AI agent systems. Traditional monitoring involves scanning CLI outputs, checking dashboard statuses, or parsing log files, which is cognitively taxing and slow to reveal bottlenecks.
2. Target Audience: AI Engineers and Developers managing multi-agent workflows (e.g., using Paperclip for autonomous coding); DevOps and Platform Teams responsible for the operational health of AI agent deployments; Research Teams and Product Managers who need a shared, understandable view of automated task progress.
3. Use Cases: Monitoring a fleet of Paperclip agents performing automated code generation and debugging; watching multiple Claude Code sessions operate concurrently; visualizing the status of a custom script orchestrating various LLM calls (via the JSON feed); displaying a real-time agent status board on a wall-mounted screen in an office (kiosk mode).

Unique Advantages

1. Differentiation: Unlike traditional APM (Application Performance Monitoring) dashboards or plain log aggregators, Cubicle uses a spatial, gamified metaphor that leverages human pattern recognition for status, making it faster to identify issues. Compared to other visual tools, its zero-dependency, read-only design and multi-platform support offer unparalleled ease of deployment and safety.
2. Key Innovation: The abstraction of AI agent states into a universally understandable "office worker" simulation. The specific technology enabling this is its lightweight, feed-based architecture that decouples the visualization from the execution platform, combined with a client-side rendering engine that uses pure CSS and JavaScript for animations without external libraries.

Frequently Asked Questions (FAQ)

1. How does Cubicle connect to AI agents like Claude Code? Cubicle does not directly connect to the AI model. For Claude Code, it uses a one-time command (npx @caglarutkuguler/cubicle install-hooks) to install lifecycle hooks into Claude Code's configuration. These hooks write status updates to a local file that Cubicle then reads via its claude-code source adapter, visualizing each session as an agent.
2. Is Cubicle secure for production monitoring? Yes, by design. Cubicle operates in a read-only mode, making only HTTP GET requests to predefined endpoints. When used with the --redact flag, it strips sensitive data (task details, commands) on the server side, showing only IDs and statuses in the UI, making it safe for shared or public displays.
3. Can I use Cubicle with my own custom AI agent scripts? Absolutely. You can integrate any system by having your script output status updates in Cubicle's simple JSON feed format (documented in FEED.md). You then point Cubicle to the URL or file path of this feed using the --source parameter for real-time visualization.
4. What are the system requirements to run Cubicle? The requirements are minimal. You need Node.js version 18 or higher. The server itself has zero npm dependencies. It runs on any system where Node.js is available, from a local laptop to a cloud server, and consumes minimal resources.
5. How does the Telegram notification integration work? When configured, the Cubicle server runs a lightweight Telegram bot. When an agent raises a "hand" (enters a waiting state), the bot sends a message to your predefined Telegram chat with the agent's question. Your reply sent via Telegram is then posted back to the agent system (e.g., as a comment on a Paperclip issue) as the required human input.