Munder Difflin provides a local desktop harness for orchestrating multiple Claude Code AI agents. It transforms terminal sessions into visual, self-coordinating teams with memory, mailboxes, and an interactive "office floor" interface, enabling users to manage and observe complex AI workflows.

How It Works

The system wraps real Claude Code terminal processes as autonomous agents, each equipped with long-term memory and a local mailbox. These agents are visualized as avatars on a 2D office floor, coordinated by a central "GOD" orchestrator agent. Communication and state are managed via an on-disk "hive" (git repo, atomic files, blackboard), with a fast, markdown-first memory layer providing semantic recall. The GOD agent routes tasks, adjudicates, and escalates critical operations to the user.

Quick Start & Requirements

• Prerequisites: macOS (primary, Windows/Linux untested), Node.js 18+, npm, C/C++ toolchain (Xcode Command Line Tools for macOS: xcode-select --install), Claude Code on PATH. Optional: semantic memory index for instant recall.
• Install & Run: Clone the repository, run npm install (may require npm rebuild for node-pty native addons), then npm run dev to launch the Electron app.
• Setup: An initial onboarding wizard guides the first-time setup.

Highlighted Details

• Real-time Agent Interaction: Full PTY control for spawned commands (defaulting to Claude), with live streaming over IPC.
• On-Disk Hive: Robust multi-agent layer featuring per-agent identity, long-term memory, atomic mailboxes, a shared blackboard, and a single-committer git repository.
• Instant Memory Recall: A markdown-first memory layer with a semantic index enables millisecond recall, degrading gracefully without the index installed.
• Visual Office Floor: An interactive Pixi.js scene displays agent avatars, pathfinding, and animated message deliveries.
• Command Center (v0.2.0): A comprehensive UI for managing agents, memory, activity (telemetry, cost), tasks (kanban board with dependencies), and schedules.
• Observability (v0.2.0): Integrates live OpenTelemetry, per-model cost attribution, fleet monitoring, context-window gauges, and a cost/runaway circuit breaker.
• Durable Persistence (v0.2.0): SQLite-backed storage for window state, history, and a durable cost ledger, surviving restarts.

Maintenance & Community

Active development is indicated by the recent v0.2.0 release, which introduced significant enhancements in observability and durability. The roadmap includes future features like chat integrations (Slack/Telegram) and pluggable agent CLIs. Contributions are welcomed via CONTRIBUTING.md.

Licensing & Compatibility

The project's source code is licensed under the MIT License. However, the included pixel art assets (tilesets, character bases) are under the LimeZu FREE VERSION license, which restricts use to non-commercial purposes only. Commercial applications will require replacing these assets or obtaining a separate paid license from LimeZu.

Limitations & Caveats

The project is primarily developed for macOS; Windows and Linux support are explicitly noted as untested. The non-commercial asset license presents a significant restriction for commercial adoption, necessitating asset replacement or licensing. The project is described as an "early prototype with a lot of surface area."