pixtuoid: AI coding agents shown as pixel-art office coworkers

pixtuoid is a terminal-based multi-agent dashboard that reimagines how developers monitor parallel coding agents. Instead of alt-tabbing between terminals, users launch pixtuoid and connect their agent CLIs (Claude Code, Codex, Antigravity, DeepSeek-Reasonix, CodeWhale, Copilot CLI, opencode, Cursor CLI, Hermes Agent, Oh My Pi, OpenClaw, Grok Build, and Kimi Code CLI) via the Sources panel. Each agent appears as an animated pixel-art character at a desk in a shared office; the office auto-expands to multiple floors as agents are added.

The visual design conveys rich status information without text. Agents type while working, display a question mark when waiting for user input, and sleep under little z's when idle. Each desk's monitor glows with the tool in use: blue for Edit, orange for Bash, cyan for Read. Shirt and pants colors derive from the agent's working directory, so colleagues from the same repository appear in matching colors. A token meter stacks paper on each desk through consumption tiers (250K, 2M, 16M tokens). Users can hover over agents for session duration and tool-call counts, or hover furniture (desks, sofas, plants, vending machines, printers) for their names. The dashboard also features per-floor office pets (a cat or dog that roams and sleeps), an OpenClaw gateway mascot (a lobster whose movement reflects gateway health), and six UI themes with day/night cycles and weather effects (rain, storm, snow, fog, overcast, wind, smog). A lofi soundtrack synthesized entirely in code adjusts to time of day, activity level, and weather, with sound effects for door chimes, printers, and vending machines.

Installation is available via Homebrew, npm, Cargo, and prebuilt binaries. Users launch pixtuoid, press s to open Sources, connect their agent CLI, then start the agent in another terminal—a character walks in from the elevator within a second. Keyboard shortcuts allow navigation (q quit, p pause, t theme picker, m sound, Tab for dashboard view, arrow keys or j/k or Page Up/Down to scroll floors, click an agent to bring its terminal to front). The tool runs entirely locally with no network connections or telemetry; agent transcripts are read only to animate the office, and session data never leaves the user's machine.

Pixtuoid is open-source under the MIT license. The codebase is organized into five Rust crates with zero terminal dependencies in the core. The architecture uses a hook shim (a 200ms fire-and-forget write to a Unix socket or Windows named pipe that cannot block the agent) and JSONL transcript watching; both feed one event channel, a reducer folds events into office state, and the renderer draws half-block pixel art. Configuration lives in ~/.config/pixtuoid/config.toml (created on first launch, all keys optional); users can override settings via CLI flags. Contributing is welcomed, especially new themes, sprite polish, and Source adapters for additional agent CLIs. The project includes full architecture diagrams, a CLAUDE.md document covering load-bearing invariants, and SECURITY.md documenting trust boundaries.

The friction of monitoring multiple AI agents in parallel has been a practical pain point for developers: each agent runs in its own terminal session, forcing constant context-switching to check status. pixtuoid addresses this by borrowing UI ideas from simulation games and desktop pets—the office metaphor is not just whimsical but functional. Each agent becomes a visual character at a desk, with animation state (typing, waiting, sleeping) and desk-level metadata (token burn, tool in use, session duration) all legible at a glance. The color-coding by working directory turns the room into an implicit org chart, and the token-meter stacks paper piles to make consumption visible.

The architecture itself is lean: a hook shim (200ms, fire-and-forget, cannot block the agent) and JSONL transcript watching feed a single event channel; a reducer folds events into office state, and a renderer draws pixel art. The tool integrates with ten agent CLIs and the OpenClaw gateway by implementing a modular Source trait, making it straightforward to add support for new tools. Security is enforced locally—hook installation, socket ownership, and transcript reading are all documented in SECURITY.md—and the dependency set is audited daily for vulnerabilities.

No comments yet. Be the first to share your thoughts!