---

[!NOTE] The world's best agents. The world's worst paper company. Munder Difflin takes the terminal-agent CLIs you already run (claude, agy, codex, grok, kimi, qwen, opencode, crush, pi, and copilot) and turns them into a self-coordinating team: each agent gets long-term memory, a mailbox, and a desk on a 2D office floor, and your clone (Michael) routes work between them while you watch. He's the boss of the floor; you're still the boss of him.

Contents

• Supported agents
• What it is
• How it works
• Features
• Getting started
• Architecture & project structure
• Roadmap
• Contributing
• Telemetry
• License
• Acknowledgements

Supported agents

Bring the CLI you already pay for. Every one of these runs as a real process in its own terminal, with your existing subscription and its hourly limits. If it runs in a terminal, it can run here.

Claude Code Codex · GPT Grok · xAI Kimi Code Gemini CLI Antigravity · Gemini Qwen OpenCode Crush · Charm Pi GitHub Copilot Cursor + any custom command

Plus bring your own keys and local models through Ollama, LM Studio or vLLM.

What it is

Munder Difflin is a desktop app that wraps real terminal-agent CLIs as fully-capable agents, wires them into a hive mind, and puts your clone in charge. Michael is the one agent you talk to in order to get things done. Under the hood it runs the fastest memory layer in the world so every agent remembers what it learns and recalls it instantly.

• Every terminal is an agent. Each claude, agy, codex, grok, kimi, qwen, opencode, crush, pi, copilot, or custom session runs as a real process in a pseudo-terminal (node-pty), byte-for-byte authentic, rendered with xterm.js.
• Every agent is an avatar. Sessions appear as characters on a Pixi.js office floor. They walk to stations as they work, and envelopes fly desk-to-desk when they message each other.
• The hive coordinates them. Agents read their memory and drain a mailbox; the router moves messages between inboxes; the GOD agent adjudicates, assigns, and escalates only when it needs you.
• Memory that's instant. A markdown-first memory layer with a semantic recall index means agents remember across sessions and recall in milliseconds.

How it works

            you ── talk to ──►  ┌─────────────┐
                                │  GOD agent  │  orchestrator / supervisor
                                │ (Michael's  │  roster · routing · adjudication
                                │   office)   │  blackboard · task ledger
                                └──────┬──────┘
                                       │ assigns · routes · escalates
              ┌────────────────────────┼────────────────────────┐
              ▼                         ▼                         ▼
        ┌───────────┐            ┌───────────┐            ┌───────────┐
        │  agent A  │  message   │  agent B  │  message   │  agent C  │
        │ provider  │ ─────────► │ provider  │ ─────────► │ provider  │
        │  + memory │            │  + memory │            │  + memory │
        └───────────┘            └───────────┘            └───────────┘
              └──────── shared hive: memory · mailbox · blackboard · log ───────┘

1. You spawn agents: each is a normal terminal process (claude, agy, codex, or custom) with its own working directory, identity, and provider-specific lifecycle.
2. Agents collaborate through the hive: a local git repo of plain files. They write to their own outbox/; the harness's router delivers into recipients' inbox/. No agent ever touches git (single-committer design avoids index.lock corruption).
3. The GOD agent runs the floor: it reads every request, resolves routine ones itself (keeping the system fully autonomous), and only escalates critical items (spend, destructive ops, scope changes) into an approvals queue you act on.
4. Everything is visible: you watch avatars move, envelopes fly, and the live terminal stream; you can type back into any session, browse its files, and read its git history.

See HIVE.md for the full multi-agent design, SPEC.md for the terminal/event plane, and DESIGN.md for the visual system.

Features

The floor

• Every terminal is a real agent. Claude Code, Antigravity (Gemini), OpenAI Codex, xAI Grok, Kimi Code, Gemini CLI, Qwen, OpenCode, Crush, pi.dev, GitHub Copilot CLI, Cursor, or a custom command, each in its own node-pty PTY, rendered with xterm.js.
• Every agent is an avatar. A Pixi.js office floor where agents walk to stations, envelopes fly desk to desk, and avatar state reflects real work.
• A GOD orchestrator you talk to. It routes tasks, adjudicates traffic, and escalates only what needs a human. Or press Talk and run the floor by voice.
• Per-agent git worktrees. Optional isolation so parallel agents never collide on branches.
• More than one floor (0.5.3). File, New Floor opens another office on its own folder, in its own window, with its own agents, board and memory. A folder only ever opens in one floor, and one licence covers them all.

Memory & coordination

• The hive: per-agent memory, atomic-file mailboxes, a shared blackboard, an append-only event log, single-committer git.
• Semantic recall: markdown memory mined into a shared palace, searchable from the UI, with condensation so it doesn't grow forever.
• Enterprise Knowledge Graph: your own documents and policies, queryable by any agent.

Control & safety

• Human gates: spend, scope, and destructive ops escalate to you. Steer mid-run or stop gracefully.
• Circuit breaker: a steer → constrain → stop ladder for agents that loop, storm errors, or blow their budget.
• Budgets & telemetry: per-agent token budgets, real cost from transcripts, a durable ledger, OTel spans, and a tool waterfall.

Command Center

• Tickets with keys (0.5.3): every card gets a key like V53-299, each agent has a Tasks tab with what it did and when, and the Tasks screen filters by date.
• Kanban tasks with dependencies, scheduled missions + heartbeat, live fleet monitoring, memory search, activity log, and a CI watcher.
• Skills: what every agent can already do across Claude Code, OpenCode and Codex, plus a browsable catalog of 227 more with search, filters, install and uninstall.
• Built-in Monaco IDE: file tree, editor tabs, save, plus CHANGES · HISTORY · COMPARE git rails with commit graph, diffs, branch compare, and guarded checkout. All fs/git access brokered through main.

Getting work in and out

• Slack & webhooks: message a channel or POST a webhook; Michael can spawn an ephemeral worker, reply in-thread, and tear it down.
• Shareable hires + Agent Gallery: import a role from a munderdifflin://hire link; import only pre-fills the form, a human still spawns it. Browse roles at the Agent Gallery.
• Local transcription and dictation (0.5.3): dictate into any app (Option on the Mac, Control+Alt+Space on Windows and Linux X11), and Stapler in Pro transcribes meetings with You and Them on Mac, Windows and Linux. A bundled Whisper model is the default engine; Apple's on device recogniser on macOS 26 and Groq with your own key are options. A built in vocabulary of 126 technical names keeps Anthropic, Claude, Kubernetes and ARR spelled right, and you can add your own.
• Tell an agent from the Stapler (Pro): Record message, several screenshots per message, PDFs, videos and folders, and a To menu that remembers who you picked.
• BYOK keys + local LLMs: per-provider keys in a write-only secret broker, plus Ollama / LM Studio / vLLM base URLs. Guides: open models · Mac Mini.
• Updates in one click: the title-bar badge runs the real update. It downloads the build for your machine, then restarts and installs it, and it reads latest once a check confirms you are current. A manual download is the fallback for when the updater cannot fetch the build itself. The first run afterwards opens that release's notes as a designed page rather than a version number.
• Your language: English, Simplified Chinese and Arabic, with right to left layout for Arabic. English is the default and nothing changes until you pick another one in Settings. The app does not read your OS locale. All three app fonts ship inside the bundle, so nothing is fetched at boot.
• Prerequisites: one Settings page showing which supporting tools (uv, git, Node, MemPalace, each agent CLI) you have, what each is for, and a button that asks Michael to install what is missing.

[!NOTE] Status: v0.5.3. Cancel your Granola and Wispr Flow subscriptions. The Stapler does both, all running locally, and every agent gets easier to start, restart and steer. Dictate into any app: hold Option on the Mac (Microphone and Accessibility permissions), or Control+Alt+Space on Windows and Linux X11, talk, and let go. Wayland says why it cannot. Meetings hear both sides in Stapler, part of Pro: You and Them, on Mac, Windows and Linux, transcribed locally, then corrected, described and sent to any of your agents. Transcription on your machine: a bundled Whisper model by default, an optional 190 MB larger model, Apple's on device recogniser on macOS 26, and Groq only if you add a key. 0.5.3 transcribes English. The bundled model is English only; on macOS 26 dictation runs on Apple's engine, which follows the Mac's language and we have tested it in English. Other languages arrive as a model download in a later release. A missing CLI is a card, not an error: install and sign in from inside the app for every engine, with Set up manually when either fails. Ask me on every agent's Inbox, answered from one box over the composer. More than one floor: File, New Floor (Shift+Command+N, or Control+Shift+N) runs another office on its own folder, with its own agents, board and memory, and one licence covers every floor. Send now goes straight in, Restart & Continue keeps the conversation for every engine, and Opus 5.5 is the default Claude model. Settings, redesigned: one Save for everything, inbound webhooks for GitHub, Linear, Telegram and your own, each with the agent that answers it, and custom secrets stored encrypted beside your provider keys. Crashes keep your work: a crashed agent keeps its uncommitted work. The IDE opens on Command+I with find and replace, the sidebar's reorder handle sits on the right in Arabic, and code in the IDE reads left to right in every language. Automatic compaction runs every 40 minutes at 30 percent of the context window, the same bar on every model size, where the previous default was every two hours at 60 percent. Tasks you can follow: every ticket gets a three letter key like V53-299, taken from the hive folder's name or set in Settings, Autonomy. Each agent has a Tasks tab beside Inbox and Terminal with what it did and when, the Tasks screen filters by Today, This week, This month or All time, and a card drops anywhere in a column. Hover a ticket in the sidebar for its details. The Pro sidebar is as wide as you drag it, and one setting shows only agents and notes. With more than one floor open, each window and Dock icon carries the office's number. Stapler fixes: the mic stops at the click, no words are lost at a pause, Record message never gets stuck, the drag follows your pointer, Reset puts it in the middle of the screen, and it never ends up off screen. Fixes: a restarted agent's terminal takes the mouse again, the hive repository no longer swallows agent git folders into its object database and the disk space comes back, and one hook event now renders one sidebar row with the task ledger read once. Not in this release: the Kimi scroll and queue bug, which we could not reproduce, and it is first on the 0.5.4 list. Launch offer: the Pro annual plan for $150 USD, adjusted for purchasing power in different countries and as low as $100 a year. Get Pro Before that, v0.4.6 brought a Simplified Chinese and Arabic interface with right to left support and self hosted fonts, the update badge that runs the real download and restart, and the security work on CLI name validation and analytics. If you're on 0.3.8, update: that build's usage-limit guard never released the agents it held, and it has been removed entirely. macOS (signed & notarized), Windows, and Linux builds are on the download page. The open source build is on the releases page.

Getting started

Download the app

Most people want this one. Munder Difflin is free and open source. Signed and notarized macOS builds of 0.5.3, plus Windows and Linux, are on the download page. The open source build is on the latest GitHub release. Install it, open it, and the wizard takes you the rest of the way. You do not need Node, a toolchain, or this repository.

You do still need at least one agent CLI on your machine. From 0.5.3 a missing CLI shows as a card in the agent's terminal: Install runs the command, you sign in inside the app, and the agent starts. Settings → Prerequisites lists the rest.

Pro adds Stapler and the new sidebar. Launch offer: the annual plan for $150 USD, adjusted for purchasing power, as low as $100 a year. Get Pro

Build from source

Everything below is for contributors and for people who want to run an unreleased build.

Prerequisites

• macOS, Windows, or Linux.
• Node.js 18+ and npm.
• A C/C++ toolchain for node-pty's native addon. On macOS, install Xcode Command Line Tools:

  xcode-select --install
  

• At least one supported agent CLI on your PATH, one of Claude Code (claude, the default), Antigravity (agy), OpenAI Codex (codex), xAI Grok (grok), Kimi Code (kimi), Gemini CLI (gemini), Qwen (qwen), OpenCode (opencode), Crush (crush), pi.dev (pi), GitHub Copilot (copilot), or Cursor (cursor-agent). Most missing CLIs self-heal: the harness runs the installer in the terminal and continues into the new binary.
• Optional: your own API keys and local LLMs in Settings → AI Engines (Ollama / LM Studio / vLLM).
• Optional: the semantic memory index for instant cross-session recall. Markdown memory works without it.

Install & run

git clone https://github.com/chaitanyagiri/munder-difflin.git
cd munder-difflin
npm install        # postinstall rebuilds node-pty against Electron's ABI
npm run dev        # launches the Electron app with hot reload

On first launch you'll go through the onboarding wizard, then land on the floor. Use Add agent to spawn your first session. The GOD agent seats itself in Michael's office automatically.

Other scripts

npm run build      # production build via electron-vite
npm run preview    # preview the production build
npm run typecheck  # type-check the node (main/preload) and web (renderer) projects

If node-pty fails to load after an Electron upgrade, re-run npm install (the postinstall hook runs electron-rebuild against the current Electron ABI).

Architecture

Two data planes feed one renderer: a terminal plane that owns the PTYs, the filesystem and git, and an event plane that runs the hive, the hook server and the router. The renderer talks to both only through a typed bridge.

The full diagrams, the module by module project structure, and the design system live in docs/ARCHITECTURE.md. They moved out of this file so it can explain the product rather than the codebase. Also see HIVE.md for the multi-agent design, SPEC.md for the terminal and event plane, and DESIGN.md for the visual system.

Roadmap

Shipped through v0.5.3: local dictation into any app and meetings with both sides of the call, with the model in the installer, the new agent cards in the free build and a new Pro sidebar, a missing CLI you install and sign in to from inside the app, Ask me on every Inbox, more than one floor on one machine, ticket keys and a Tasks tab on every agent, redesigned Settings with inbound webhooks and custom secrets, Opus 5.5 as the default Claude model, one compaction bar for every model size, a Simplified Chinese and Arabic interface with right to left support and self-hosted fonts, twelve agent engines with BYOK keys and local LLMs, voice orchestration, the hive (memory · mailboxes · blackboard · event log), Command Center with kanban and weekday schedules, a built-in Monaco IDE with git rails, integrations registry + secret broker, Slack-spawned workers, shareable hires and the Agent Gallery, observability and the circuit breaker, durable persistence, session resume, multi-window floors, one click updates, a Skills browser, a live Prerequisites check, cost reporting folded from the ledger, semantic memory that works on Apple Silicon, and an updater that installs the build instead of pointing at it. Full history in CHANGELOG.md.

Next up:

• More chat integrations: Telegram and richer chat bridges that pipe a channel into Michael's queue and route replies back out.
• More engines & integration templates: keep growing the engine roster and the integrations registry.
• Fuller avatar coverage: drive the remaining station visits and tool-bubbles entirely from real hook events.
• Durable layout & command history: extend persistence to agent layout and per-session history.

Contributing

Contributions are welcome. This is pre-release software with a lot of surface area. Start with CONTRIBUTING.md. The short version: fork, npm install && npm run dev, keep npm run typecheck green, and derive any new UI from DESIGN.md tokens. Good first areas: wiring real hook events, the add-agent flow, the config drawer, and cross-platform work.

[!IMPORTANT] Every pull request must show a before and an after: screenshots, or a recording when the thing moves, under the ### Before and ### After headings in the PR template. This is checked automatically and a PR without it does not merge. "My change has no UI" is not an exemption; it just changes what the evidence looks like. See Evidence is mandatory.

Questions, bugs, or want to show off your office? Join the Discord: https://discord.gg/SEDzP5ZPk5. Add your Discord handle to a PR and you'll get the employee of the month role when it merges.

Looking for somewhere to start? The good first issue list is kept stocked with small, self contained work that has a clear finish line.

Everyone whose code is in Munder Difflin is listed in CONTRIBUTORS.md. If that is you, it is yours to point at. The list is generated from the pull requests themselves and updates on its own, so you appear without having to ask. It also names the contributions that are in main but that GitHub shows as closed rather than merged, because that was our mistake to record and not theirs to explain.

Telemetry

Official builds send a small set of anonymous usage events (app opened, agent spawned, feature used), never prompts, code, file paths, or agent output. The complete event list, the anonymity guarantees, and the three ways to opt out (Settings toggle, DO_NOT_TRACK, or building from source, forks compile with no key and send nothing) are documented in TELEMETRY.md.

License

[!IMPORTANT] Asset licensing. The bundled pixel art (tilesets and maps) is Modern Interiors - RPG Tileset [16X16] by LimeZu, used under the Complete Version licence, which permits editing and use in commercial and non-commercial projects. Credit to LimeZu is required by that licence and must stay in place. The Office cast is not LimeZu art. It is drawn procedurally in portraitArt.ts. See src/renderer/src/assets/ATTRIBUTION.md.

The source code is licensed under the MIT License. See LICENSE. The MIT grant covers the code only; the bundled pixel art is licensed separately from LimeZu and is carved out in LICENSE-ASSETS. Munder Difflin is an affectionate parody and is not affiliated with NBC's The Office or Dunder Mifflin.

Acknowledgements

• LimeZu for the Modern Interiors pixel-art tilesets (Complete Version licence).
• shahar061/the-office for the office tileset/map vendoring.
• Pixi.js · xterm.js · node-pty · electron-vite · CodeMirror for the libraries this is built on.
• Remotion for the landing page's animated "how it works" clips (landing-remotion/).
• The Office (US) for Munder Difflin, Inc.