Munder Difflin – Multi-Agent Office Floor
An open-source desktop harness that turns the terminal coding CLIs you already use — Claude Code, Codex, Gemini CLI, etc. — into a self-coordinating team managed by your own clone.
Evidence shows: README and SECURITY.md describe a local-first architecture, IPC via contextBridge, path validation, single-committer git, and human approval gates (spend, destructive ops). But these are claims without code verification. Deductions: lack of actual code evidence for least privilege, user confirmation mechanisms not shown in code, data flow transparency only documented, sensitive data handling (e.g., keys) not detailed, dependency security not audited, external effects (e.g., network) not explicit, rollback mechanisms not described, source attribution only based on author claims.
Evidence shows: README and CHANGELOG describe features, but self-consistency is moderate due to some inconsistencies (e.g., version numbers, feature descriptions). Dependency availability not verified, failure messages not shown in code. Deductions: dependency availability not verified, failure messages not shown in code.
Evidence shows: README describes multiple scenarios (multi-agent, different CLIs, local models), audience is clear (developers, teams). Capability boundaries are described (e.g., approval gates, circuit breaker), but trigger precision not detailed, environment fit described (macOS/Windows/Linux). Deductions: capability boundaries and trigger precision lack concrete code or configuration examples.
Evidence shows: README structure is clear, install notes detailed, license explicit (MIT), CHANGELOG and version present. But naming stability not verified, examples and FAQ limited, known limitations not explicitly listed, maintenance responsibility not clear. Deductions: naming stability, examples, known limitations, maintenance responsibility insufficient.
Evidence shows: README describes outputs (e.g., terminal streams, memory search), marginal value high (multi-agent coordination), cost-benefit reasonable (uses existing subscriptions). But output usability not verified, cost-benefit not quantified. Deductions: output usability not verified, cost-benefit not quantified.
Evidence shows: README and CHANGELOG provide some claims, but lack traceability, cross-source verification limited, facts and inferences not clearly separated. Deductions: claims lack code or test support, cross-source verification insufficient, facts and inferences not clearly separated.
- Static review cannot verify security claims such as path validation and approval gates.
- Dependencies like localtunnel and tunnelmole may introduce network exposure; review needed.
- Publisher identity is unverified; treat with caution.
What does this agent do, and when should you use it?
Munder Difflin is a local-first, multi-agent orchestration harness (desktop app) that packages real terminal AI coding CLIs such as Claude Code, Codex, Gemini CLI, and Grok into fully capable agents. Each agent gets long-term markdown memory, a mailbox, and a desk on a 2D office floor where it appears as an avatar. A central GOD agent (named Michael) acts as your clone: you brief him, he routes tasks, resolves routine work, and escalates only critical items to you. Underneath, a 'hive' — a local git store of plain files — provides shared memory, atomic-file mailboxes, a blackboard, and an append-only event log. The app is built with Electron, React, TypeScript, uses Pixi.js for visualization and node-pty/xterm.js for real terminal processes. It is free, open source (MIT), and available for macOS, Windows, and Linux with signed/notarized macOS builds.
Munder Difflin wraps a dozen terminal-agent CLIs—including claude, codex, grok, kimi, gemini, qwen, opencode, crush, pi, and copilot—by launching each as a real process inside its own node-pty pseudo-terminal, rendered byte-for-byte with xterm.js. It creates a 'hive' folder that is a local git repository containing per-agent markdown memory, atomic-file mailboxes, a shared blackboard, and an append-only event log (single-committer to avoid lock corruption). The GOD agent (Michael) monitors all requests, resolves routine ones autonomously, and only escalates spend, destructive actions, or scope changes to a human approval queue. Agents communicate by writing to their outbox/; the router moves messages into recipient inboxes. The UI displays a Pixi.js office floor where agents walk to stations and envelopes fly; you can click any desk to view the live terminal, type back into the session, and inspect files or git history. It also includes a Command Center with kanban tasks, scheduled missions, memory search, a Monaco-based IDE with git rails, a skills browser, Slack/webhook integrations, and an in-app updater.
- A solo developer who wants to run several parallel Claude Code sessions on one machine with a central coordinator that retains context.
- A team already paying for Claude Code or Codex subscriptions that wants to automate multi-step work without paying extra API fees, sharing the hourly limits across agents.
- A developer who wants a visual overview of what multiple agents are doing, with the ability to steer or stop any session from a live terminal view.
- A user who wants to import ready-made roles from the Agent Gallery (via munderdifflin://hire links) and quickly spawn an agent with a predefined identity.
- An operator who wants to trigger work from Slack or a webhook, letting Michael spawn an ephemeral worker, reply in-thread, and tear it down afterwards.
- A privacy-conscious developer who prefers local processing, wants to bring their own keys, or connect to local LLMs via Ollama, LM Studio, or vLLM.
What are this agent's strengths and limitations?
- Leverages existing subscriptions (Claude Code, Codex, etc.) — no additional model fees, within each provider's hourly limits.
- Built-in shared, semantic memory across sessions via a local markdown hive and memory palace, so agents remember what they learn.
- Visual office floor with live terminals, memory search, and activity log gives transparency into what each agent is doing.
- Designed for controlled autonomy: per-agent autonomy levels, human approval gates for spend/destructive operations, and a circuit breaker that can stop runaway agents.
- Still pre-release (v0.4.6) — expectations should account for rough edges; the 0.3.8 build had a known use-limit guard bug.
- Requires at least one supported terminal AI CLI to be installed; node-pty native compilation needs a C/C++ toolchain when building from source.
- Tied to specific commercial CLI subscriptions or BYOK keys (though local models are supported), so you cannot run it fully free or fully offline without those.
- Desktop Electron app with many subprocesses can be resource-heavy; no official CLI/headless mode documented.
How do you install or deploy this agent?
For most users: download the signed and notarized macOS, Windows, or Linux build from the releases page (https://github.com/chaitanyagiri/munder-difflin/releases/latest) and install it — no Node.js or source required. To build from source: prerequisite are Node.js 18+ and npm, plus a C/C++ toolchain for node-pty. Then run: git clone https://github.com/chaitanyagiri/munder-difflin.git && cd munder-difflin && npm install && npm run dev. You must have at least one supported agent CLI (e.g., claude for Claude Code) on your PATH; the app can install missing CLIs for you from Settings → Prerequisites.
How do you use this agent?
On first launch, the onboarding wizard checks for required tools and offers to install anything missing. Then click 'Add agent', pick your CLI, model, and autonomy level, and give it a desk — the GOD agent seats itself automatically. Brief Michael (the coordinator) with your task; he will assign and route it. Click any agent's desk to view and interact with its live terminal, or browse its files and git history. Use the Command Center to view kanban tasks, schedule missions, and monitor activity. For a development build, run npm run dev after installing dependencies.