v0.2.6 · Multi-provider · Open source · macOS · Windows

Run a virtual office of AI agents on any computer.

Munder Difflin is a highly capable local multi-agent harness that makes a multi-CLI agent cluster work 24/7 for you. It supports Claude Code, Codex and Antigravity CLI, and comes with Slack and webhook integration. Oh — and all the agents look like The Office characters.

Free & open source. macOS · Windows · Linux — or build from source in two commands.

munder-difflin — the floor
Munder Difflin running a floor of Claude Code agents — avatars at their desks, the GOD orchestrator in Michael's office, and a live terminal session.
What you get

Everything to run your office. Out of the box.

🔌
Supports Codex, Claude Code & Antigravity CLI
Three major AI CLIs, one unified harness. Each gets a desk, a mailbox, and full hive integration.
No setup required
Download, open, and your office floor is live. No config files, no YAML, no wiring.
💬
Slack & custom webhook
Trigger and monitor work from Slack or any POST-capable tool. The GOD orchestrator picks it up and replies when done.
📅
Schedules, monitoring & token budget
Recurring missions, live fleet status, real-time cost tracking, and per-agent token budgets — all in one place.
🧠
Memory layer
MemPalace gives every agent searchable long-term memory. What the office learns once, it never forgets.
🐝
Hive mind
All agents share one orchestrator, one memory, and one task board — so the whole floor moves as one organism.
📨
Inter-agent communication
Agents message each other directly — routing subtasks, sharing findings, and coordinating without you in the loop.
Runs for days
The harness is built for long-horizon work. Set ambitious goals, step away, and check back when it's done.
🤖
Automate anything in one prompt
Describe a complex workflow in plain language and the GOD orchestrator builds it for you. For example: we built a custom CodeRabbit-like PR reviewer for this repo — automated code review on every PR — in a single prompt.
The why

Yes, it's fun to watch.
It's also genuinely useful.

Beyond the joy of watching a floor of agents build your thing and chat around, four reasons a local hive does more than a lone terminal ever could.

⚖️

Do more with less

A GOD orchestrator on a lean model — say Sonnet 4.6 — routes work to specialists running heavier ones like Opus 4.8. The right capability on the right task, not one model paying for everything.

🖥️

It's your harness

The harness lives on your computer. You decide exactly what each agent does, which tools it gets, and how far it can go — full control, nothing in someone else's cloud.

🧠

Memory, built in

MemPalace is wired into every agent and the shared hive mind — the best-performing memory layer we know of. Nothing the office learns ever gets learned twice.

🚀

High-agency outcomes

Agents run autonomously for hours, even days — reiterating on an ambitious project until it's actually done, not just attempted.

📋

The Command Center

Rebuilt in v0.2.0 — a dependency-aware kanban, live fleet monitoring, and per-agent budgets in one place. Assign tasks, track todo/doing/blocked/done, and wire dependencies so nothing starts before its upstream is ready.

Missions on a schedule

Set recurring missions — "triage new issues every morning", "summarise the week every Friday" — and the harness dispatches them automatically from a dedicated Schedules tab. Set it once, never think about it again.

🔭

Observability, end to end

v0.2.0 adds live OTel telemetry, per-model cost, a live fleet grid and a per-agent tool-span waterfall. Set per-agent token budgets, watch real spend in real time, and a context-window gauge on every agent card shows how full each session is.

🔔

Stay in the loop, hands-off

Native desktop notifications fire when an agent finishes a task or needs your input. CI failures, GitHub issues, and human-approval requests surface in the same Command Center without you polling.

💬

Trigger from Slack or a webhook

Kick off work from anywhere. Drop a message in Slack — or POST to a secure, opt-in webhook — and the GOD orchestrator picks it up as a task, runs it, and replies back in the thread when it's done. Your hive, reachable from the tools you already use.

💾

Survives a restart

v0.2.0 moves state into a durable SQLite store, so window layout, history, and the cost ledger persist. Quit and reopen, then one-click Restore team brings your last session's workers back to their desks.

🛟

Guardrails that hold

A cost/runaway circuit breaker and a scheduler heartbeat keep the floor stable. Step in any time with a human-in-the-loop gate, steer an agent mid-run, or stop it gracefully — all without killing the session.

📁

Your hive, your folder

Point the hive and memory home at any folder you like, with a safe move that relocates existing state. Keep your office wherever it makes sense — a project drive, an external disk, a synced path.

The how

One message in. A floor that ships for days.

You talk to a single GOD orchestrator. It routes your request to a hive of role-based agents that run in parallel — real terminals, one shared memory — on an office floor you can watch. Then they keep going, autonomously, for hours or days.

munder-difflin · how it works the full loop
TRIGGERS · KICK IT OFF FROM ANYWHERE YOU Slack Webhook Schedule THE OFFICE FLOOR GOD orchestrator reads · plans · routes routes the work research build review git · each agent in its own isolated local worktree MemPalace shared long-term memory ⟳ autonomous · hours → days
STEP 01

You talk to GOD

Describe the goal to one GOD orchestrator in plain language — like briefing a lead, not wiring a pipeline.

one chat
STEP 02

It routes the work

GOD plans the job and hands each piece to the right role agent — research, build, review, test.

orchestratorhive mind
STEP 03

Agents run in parallel

Each is a real Claude Code, Antigravity, or Codex terminal in its own pseudo-tty, all working at once — sharing one long-term memory.

node-ptyxterm.jsMemPalace
STEP 04

You watch the floor

Every agent is an avatar on a live office floor, now with new states for compacting and looping. Open any desk to read its terminal, files or git — or type back in.

Pixi.js floorlive terminals
STEP 05

It ships for days

The hive keeps iterating on its own — and you stay in control with a human-in-the-loop gate, mid-run steer, and a graceful stop, escalating spend, scope or destructive ops for a yes/no.

autonomoushuman-in-the-loop
v0.2.4 · Claude Code · Antigravity · Codex

Three CLIs. One hive. Full parity.

The floor is no longer Claude-only — and in v0.2.4, it's fully equal. Claude Code, Antigravity (Gemini · agy), and OpenAI Codex each have a lifecycle-hook bridge, a mailbox, and a seat at the same table. One orchestrator. Every function your business runs on.

💻
Engineering
code · review · test · ship
📣
Marketing
campaigns · copy · SEO content
🎯
Sales
lead research · outreach · follow-ups
📊
Analysis
reports · data · competitive intel
📝
Product
specs · roadmaps · release notes
🔍
Research
markets · users · literature
✍️
Writing
docs · blogs · proposals
⚙️
Operations
workflows · triage · monitoring

🔌 Claude, Gemini, Codex — same bridge, full parity

v0.2.4's headline: Codex now has a full lifecycle-hook bridge — the same integration Antigravity has had since v0.2.3. All three CLIs run through one unified dispatch path: live status, inbox drain, and outbox routing work identically whether the agent is Claude Code, agy, or codex. Every provider gets a desk, a mailbox, and shared memory.

🧠 Runs your whole company, not just your codebase

Spawn a sales agent to research leads while an engineering agent ships the feature, a marketing agent drafts the launch post, and a product agent writes the release notes. One orchestrator. One office. Every function, in parallel.

📡 Remote-control the whole office

Run /remote-control on your GOD agent and reach the entire floor from anywhere — check in, redirect work, clear approvals — all through the one agent that runs everything else.

you/remote-control
→ GOD online · 4 agents · 1 awaiting approval
→ "approve the deploy, pause agent.research"
done — from your phone.
Open source

Built in the open, on purpose.

We strongly believe that open-source software has been the building block of nearly every prominent technology driving this world. Building a project like this in the open is one of the best ways I know to give back to the community that made it possible. The code is MIT-licensed — fork it, learn from it, ship with it.
Get it free

Download, or build from source.

Grab a build for macOS, Windows, or Linux, or run it yourself. You'll need Node 18+, a C/C++ toolchain (for node-pty), and Claude Code, Antigravity (agy), or Codex on your PATH.

# clone & run
git clone https://github.com/chaitanyagiri/munder-difflin.git
cd munder-difflin
npm install # rebuilds node-pty for Electron
npm run dev # launches the app with hot reload
Show some support

Keep the office running.

The download is, and always will be, free. If Munder Difflin is useful to you, two small things keep it going.

Drop a star on GitHub

It costs nothing, takes two seconds, and is the single biggest way to help the project reach more people.

★ Star the repo
💛

Become a patron

Chip in to fund development and assets. Patrons get listed below with a backlink to wherever you'd like.

Become a patron →
v0.2.6 · Free & open source

Your virtual office is one download away.

A full hive of AI agents — Claude Code, Antigravity, and Codex — running 24/7, for you. Free, local, and MIT-licensed. Download for macOS, Windows, or Linux — or clone and run in two commands.

macOS · Windows · Linux · MIT-licensed · Claude Code · Antigravity (Gemini) · OpenAI Codex