Framer AI AgentsDesign and publish professional sites with AI
[Open-Source] Local Multi Agent Harness that wraps around coding agents you already pay for like Claude Code and Codex to run an office of forever running agents working for you 24/7 in "the office" styled simulation. Be the boss of this office or let your clone be the boss when you are not available. For Developers, Product Managers, Designers, Founders, Sales, Marketing, Legal, HR or anyone who works in tech.
Free
Launch Team / Built With
Munder Difflin
Hey Product Hunt 👋
It’s 2026 and building a digital clone of yourself to do all your work is possible with a claude subscription.
So I built Munder Difflin: A local multi-agent harness that uses your existing coding agents like claude code and codex to run in an office of persistent agents that does your work. Decisions run through you or a clone of you that orchestrates.
If whatever’s written above does not make sense:
It’s a free and open source PC app that runs an office of agents that can do literally anything you do on your own computer with your context, you get to see them work in a “the office” themed simulation.
Be the boss of this office yourself or let a clone be the boss when you are not available.
I have made it open source which means it is available for the whole world for free to use, it runs locally which means no personal data goes anywhere, it’s ad-free and lastly it means I did it for the love of the game.
We started from v0.0.1 about 40 releases and 2 months later we are at v0.4.1
Over 2000 people use it already, along with 677 github stars and we are finally live on Product Hunt help please upvote and share about Munder Difflin to help us get to #1 on PH today.
Find us on github: https://github.com/chaitanyagiri/munder-difflin
Try Munder Difflin now: https://munderdiffl.in
[No setup needed if you use claude code, codex or any cli agent]
Two things people always ask:
- "Isn't that a fortune in tokens?" I run it on a $100/mo Claude plan —
one Opus orchestrator, nine Sonnet workers. Shipped continuously for weeks without hitting a limit. The cheap models do the work; the expensive one decides who does what. You can use a mix of agents from different providers to run it way cheaply.
[Simulation is deterministic, they do not consume token]
- "What's the catch?" There isn't one. MIT licensed, free forever, runs
entirely on your machine. Your code never touches my servers because I
don't have any.
I'd genuinely love your feedback, especially on what breaks. I'm here all
day.
The token answer makes sense on a subscription. The part I would want to know before letting this run while I am not watching is how Michael (as I understood it, that is what my clone is called) decides a worker is finished.
I measured one day of my own agent work and the tracker put 64.9% of the spend in its abandoned bucket. None of it looked abandoned while it was running. The sessions that went nowhere looked like the ones that worked, right up until I read the output.
You say the expensive model decides who does what and that a clone can be the boss when you are not available. So what happens when a worker comes back without an answer, either asking for something or saying it cannot proceed? Does that count as done or does he tell it apart from a real result?
Munder Difflin
@dimhold ideally if not possible through a lower quality model agent the clone will orchestrate it to a higher quality agent. It's recommended to run a cluster of agents with mixed quality.
Munder Difflin
@dimhold It's advised to run agents of different capacities and qualities to run together if a less capable agent can not do something it'll either assign a more capable agent or just flag it to wait for human in loop review.
Checkout this screenshot:
I am about to lose my mind. I accidentally selected Claude Code in the initial setup, and I can't figure out how to change that (since I don't have CC access). I've uninstalled twice. #frustrated
Munder Difflin
@jckiker Hey you can select the agent type from monitor section:
@chaitanya_giri after scrubbing every folder available on my computer, I got it to re-initiate the "new install" behavior. However, when I tried selecting Codex, it just didn't work, without much feedback. At least with Claude I was able to log-in to then realize I didn't actually have access to Claude Code. I'm going to give it one more try, but then need to be more productive. Thanks for the reply!
Munder Difflin
@jckiker I am using it with 4 agents Claude Code, Codex, Antigravity and Grok Build and it works, I'll get down to the bottom of it. It'll be automatically fixed and updated in next version rest assured.
Munder Difflin
@jckiker If you want to restart you can do that too from settings by selecting a different path for your clone home.
Interesting idea, but gamification in such serious projects is a bit confusing.
How complex are the tasks your service can handle? For example, if I need a tender management system for a company, could it develop one from scratch and make it comparable to well-known off-the-shelf solutions?
Munder Difflin
@natalia_iankovych I'll answer your question in 2 parts:
1. The simulation is deterministic and does not consume tokens, there's a full screen mode that allows you do do everything in a distraction free environment. The sidebar can also be hidden.
2. It'll build a tender management system end to end, if ran with claude code with opus and/or codex with sol or terra or that level of agents. It can also alternatively set up something opensource from scratch for you aswell if you need. Discuss with it and it'll 100% get your job done end to end.
Munder Difflin
Naming it after Dunder Mifflin is doing more marketing work than most taglines 😄 congrats Chaitanya. When your clone is the boss overnight, what is it not allowed to decide?
Munder Difflin
I tested Munder Difflin v0.4.3 with Codex on Windows. The architecture around CLI workers, worktrees, mailboxes, lifecycle tracking and observability is impressive.
I hit a Windows/Codex startup issue that prevented my three-agent benchmark from running, so I’m not ready to adopt the runtime yet. But the source architecture was valuable enough that I reused several patterns in my own orchestration system.
Looking forward to testing it again after the Codex integration matures.
Munder Difflin
@muracan Hey thanks for the feedback will fix this in next release. ASAP thankyou so much