YC Just Open-Sourced the AI That Runs Their Whole Company (QM)

Y Combinator just open sourced the AI system that runs their entire company. Accounting, legal, events, engineering, all of it runs through one agent harness called QM. And as of this week, you can deploy the exact same thing for free. So I cloned it and forked it, deployed it. I read the source code and I went through their own security file line by line. And there are things there that almost nobody covering this is telling you. Now, this is the second time

In the last two weeks that a serious player has dropped something they're calling a multiplayer agent harness. If you saw the last video I did on Buzz, you already know why that phrase matters. Now, by the end of this one, you'll know exactly what QM is, whether it's actually good, and the one thing that YC admits about it that should make you slow down before you run it on anything real. Now, on July 31st, no press release, no keynote, just a

Simple Y combinator tweet saying, "We've been running an internal multi- aent harness. We call it QM." And we're putting the whole thing under an MIT license, cloud first, lives in Slack, and on the web. Now, QM is short for quartermaster. In an army, the quartermaster is the person who provisions everyone, supplies, equipment, whatever each unit needs to actually do their job. Now, that name is the whole pitch here. This isn't one shared chatbot. It's a system that provisions a personal AI

Agent for every single person in a company. And people noticed this fast. The repo went up. It's been about 9 days and it's sitting at over 12,000 stars. The launch tweet is at 2.7 million views and the hacker news thread crossed 678 points. For a brand new repo, this is a serious launch. Here's the context that makes this interesting. We've had terminal coding agents for a while now. This is things like Claude Code, Codeex. We've also had personal assistant harnesses, things

Like OpenClaw, Hermes agent. What we didn't have until about 10 days ago was anything built from day one for a whole company at once. At least something that's out there in the open yet. And QM is planting a flag in that third category. But a flag is easy to plant. The question is whether the thing behind it actually works. So, let me show you what it really is. So, imagine you give one agent to 20 people at your company. Two of

Them tell it to edit the same files at the same time. One person's scheduled job quietly overrides another person's memory. Somebody's leaked API key is now the key that every single request runs with for everybody. This is not a hypothetical. This is what actually happens when you take a personal assistant design and staple it to a team. And this is literally YC's own origin story. They started with a simple internal agent. Then it grew into more than 50 separate Hermes agents,

One per person, and that became unmanageable. M is what they built to dig out of that hole. The fix is one concept, the scope. So every person gets their own scope. Inside that scope is their files, their memory, their credentials, their scheduled jobs, and their own durable sandbox, which is basically a persistent computer where anything the agent installs actually sticks around. Channels, teams, and projects get their own scopes, too. And here's the elegant part. your memory, your files, your keys, and

Your sandbox are all wrapped in exactly one boundary, not four separate ones. Because the moment you draw two boundary lines, you've created a gap for something to leak through. And the same identity follows you everywhere. The agent that knows your contacts in the web app is the same agent with the same memory when you mention it in a Slack channel. A community comment on HackerNews put it better than the marketing did. A hard problem in multiplayer agents was never the AI

Loop. It's scoping. Um's answer to that is genuinely good. So that's the idea on paper, but let's take a look at this thing deployed so we can see how it actually runs. Now, in order to get this running, they do have a quick start setup guide in the GitHub repo, but I'm going to be straight with you because I personally did not follow this. I didn't type any single commands in the terminal or anything like that. I simply just asked my

AI agent to get this thing set up for me, which you could do. You could point it at the official QM repo and do so, but you'll also be able to get access to this 30page plus document in our free school community, which you can access in the link down below, which will literally give your AI agent everything it needs and all the potential roadblocks that I encountered along the way um to get this set up seamlessly on the first try.

Once you get it set up, this is what it will look like. There is the web UI right here and there's also the admin section right here. Now, two honest notes because like I said, it wasn't frictionfree here on Windows specifically. The official version tells you Docker is not installed when Docker is installed and running. It's a small Unix assumption in a tool check and the way through it is to run the whole thing inside WSL. Second one, and to be

Clear, this is the official version, not the way I ran it. uh it is called a local test drive. The official tool refuses to start until you set up a container registry and push a sandbox image to it. Their own error message calls it a fly app and it defaults to flies registry. Though you can point it somewhere else if you want. Either way, you cannot skip it. And the punch line is that the image you just pushed is not the

Image that runs your agent locally. That one gets built on your machine. the cloud dependency is the paperwork, not the execution, which is a strange thing to require something advertised as a local test drive. Now, both of those I fixed in this fork. So, hold that thought and we'll come back to that. Like I said, just give that to your claud code or your codeex, ask it to set it up. Like I said, resource down below and you'll be able to

Set that up. Whether you want to set up the official repo or if you want to set up the forked version, you'll have both options. Now, here's the scope model from the last section except now it's not just that diagram. This is my agent. So, we have my files. We have my memory. We have my credentials aka the keychain. We have scheduled jobs aka crrons. We have my skills. And if a teammate logged in right now, they'd get their own copy

Of all this. And neither of us could touch the others. So, if we look at the web UI dashboard, you'll see there are projects right here. So you can go ahead create new projects and within each project you can add specific people from your team. Now over here you have your chats which will be all the different chats you have with your agent. Now you will notice a few differences from my fourth version versus the official version. So with the fourth

Version you'll also see some of the Slack chats right here. So this is the actual Slack channel my agents live in. You'll also see the regular chats here. And I also added a feature here which are rooms and we'll come to that later on. Now files is where files can be either created from the agents or uploaded or shared with you and your team. You can see I have a file right here. Now crons, this is pretty great because this is

Where you can schedule your crons, create new crrons. So you can literally just tell your agent what type of cron you want in plain English right here and it will set it up for you. You can see I set one up right here for daily AI research report. And this is just basically going to produce a concise AI research report, uh, 3 to five bullets, and send it to me every single day at 8:00 a.m. You can see the recent runs

Right now. And even if I go ahead and click run now, it's going to actually run this run right now. Okay. And boom, you can see the recent run right here. So, if I go over to my Slack and I take a look at my QM Claude codebot right here, you can see that it actually sent me a research update right here with, you know, for Monday and it went through all the different points right here. I'm not going to read

Through this, but you can see that it actually worked. Now, when you first get things set up in the admin section here, you'll see the onboarding tab. So on the original repo, you do have to provide a anthropic or open AI API key. Now if you're running this in a team production environment, that may be a bit easier, but if you're running it locally, you're probably going to want to use your cloud code or codec subscriptions. In the forked version, we

Have that set up. So it will pull the harness config that you have set up on your actual computer. Now, we'll come back to this area in a second, but this is where you can actually set up your Slack apps as well as your O apps and a bunch of other stuff. Now, keychain is very cool. This is where you can share accounts and credentials that your agent may use on your behalf. You can see connected accounts, store credentials. You can

See here I just have a test one right here, but I also have this SSH key which allows QM to actually access files locally on my actual workstation. Now, I did give it a sandbox folder environment just so it can't access everything, but this is something I set up just in case I wanted to access maybe potentially skills that I have set up on my workspace and I wanted to do specific things on my actual computer. Uh, I can do so.

And in the actual document, it explains how to do that if you do want to set that up. Now, over here in apps, this is where you can see different apps that your agent deploys. this right here. For example, I told it to create a CRM dashboard. Now, if I go over here and click on open, if I go ahead and open this, you can see we have this very simple CRM dashboard right here. Now, in the admin dashboard, you can

Pretty much see an overview of everything that's going on within the actual uh QM web UI or Slack channel. So, we can see all our Slack sessions. Here we have a tab for different judgments. So this is basically a cheat model monitoring when to stay silent or when to actually speak up. You can see the logs for that here. Now the governance tab is the or control panel and this drop down decides how much your agents are allowed to do on

Their own. So strict pauses for your approval auto screens external content before the model ever sees it. And then there's a third option sitting right there in the list called dangerous. So QM has three security postures. strict where the agent pauses for approval basically on every action. Auto which is the default and then dangerous where there's no screening and no approvals at all. Now instead of YC just saying enterprisegrade security and moving on, if we take a look at the security.md

Here, YC published a list of 14 things their own system cannot fully protect against. Let me give you the ones that matter here. So one, the command policy, the thing that's supposed to block dangerous commands is their word bypassable. So if you obiscate or encode a command, it can slip through. And two, when the agent is using a credential, that credential sits in plain text in the sandbox while it's in use. And three, the agent's browser actions sit completely outside the

Approval system. The thing watching the agent isn't watching the browser. And four, admins can read almost everything. Your transcripts, your memory, your keychain metadata with nothing but an audit log to show for it. Basically, underneath all of it is one line that the whole security model rests on. If the agent acts as you, it can do anything you can do. And that's the deal. M scopes the blast radius down to one person's permissions, which is real and valuable. But inside that

Scope, the agent has your keys. Now, that's not a bug. That is the fundamental trade of giving an AI agent real access. So, the one independent review I could actually find that ran it landed on the same place. Basically, their words are, "This is a bounded pilot for a startup with real engineering muscle and narrow internal workflows. It's not something you point at customer data or run on anything that matters. It's version 0.1. Treat it like version 0.1." That's a catch.

Now, let me show you some stuff that QM does that genuinely no one else in this category does because there are two things here that I think are actually ahead of everyone. So, first, the app that we published earlier, most agents in this space hand you a transcript. This one hands you a deployed tool your team can open at a URL. The security callback is those app links are basically bearer tokens. So, if you share one, it stays live even after

You revoke access until it expires. This is a rare capability and early implementation. Now, second, and this one is underrated, QM has a genuinely smart answer to a problem every open- source tool has. How do I customize this for my company without forking it and then being stuck on an old version forever? Well, QM's answer is a deployment directory. So your custom config, your tools, your skills all live in your repo on top of a core that stays bite forbite identical

To theirs. You get all the updates without merge hell. And they even ship a skill that syncs your private copy with upstream and pushes generic fixes back with your company's name scrubbed out first. And the whole thing here is model agnostic by design, which you know is a phrase a lot of tools use. So let me just show you here in our chat. We can simply choose claude code or codeex right here. We can set the specific model like fable opus

5 ppt 5.6 soul. Now right now I have claude code as default but if I wanted to switch to codeex I could do so right here. And then we have our effort levels. So auto low, medium, high, extra high, max, ultra code. Right. So pretty much everything I've shown you so far is Y Cominator's version as they shipped. Now, let me show you mine because like I said, I forked this whole thing and went to work on it for the last

Few days. Everything from here on is stuff that my fork does that the official version doesn't do exactly as I'm showing. So, my fork runs native on Windows, so no WSL needed. And it starts with zero cloud accounts, no registry, no fly apps, nothing. Now, you could of course still use those if you want to deploy to production with a team. And like I mentioned, it also uses claude and codec subscriptions that you're already paying for. You don't have to use

API credits. So, with regular QM, you can talk to your Slack agent that you have in your Slack channel, or you could, of course, use the web UI and talk to your agent like so. And right now, I'm using the Claude code harness with this one. I can talk to it. It's going to respond back. And really, that's that. Now, if you're familiar how Buzz works, I really like the concept of being able to have multiplayer. Yes, multiplayer people talking to

Your agents, but also having multiple different agents on different harnesses using different models to get different perspectives and feedback. And that's where I've seen really good outputs from my agents. So this is one of the main parts I built this fork for. So here if I go ahead and say to both of these agents that I have, codeex and claude code, um, which I have two different agents running on two separate hardnesses, and I say debate, should a small team run

One powerful AI agent army or many specialized ones, go back and forth two times, keep each reply to two to three sentences, and land on a recommendation. And I'm going to go ahead and send this. Now, we can see our codeex agent is starting to reply in this thread right here. If I click into it, we can see, you know, it's giving the response right now. And it's tagging the Claude code agent. Now, Claude Code is giving its first message out

Of two messages here. And now it's also tagging uh Codeex back for the second message. Okay. Now, we got Codeex's second message right here. And we can see Claude Code agreed with its second message. I'm not going to read all this, but you can see it gave its recommendation for a small team, one agent plus a handful of narrow pool scope sub agents it can call. Not a swarm appears and never more agents than you can evaluate. So each agent here

Has its own name, its own avatar, its own brain running different harnesses. Now this is a thing that I invented in the fork called rooms. Now the official version has nothing like it. This is how the agents take turns like colleagues instead of one bot just talking to itself. So there's even an agent client protocol harness in the fork. So the same protocol that Buzz is standardized on. So anything that speaks ACP can drive one of these agents as well. Now

What happens if I ask these agents to do the same thing but go back and forth 10 times. This is the thing that I mentioned in my last video about Buzz that agents that can talk to each other can essentially spend your money or tokens forever. I want to give credit where it's due because the official version of QM does ship a break dollar budget per person and per or per day, but it doesn't do anything about conversational length back and

Forth between agents because the official version doesn't have multiple agents. It just has one agent. So since we added in the multi- aent conversation, we added in a break as well. So the budget has layers. So there's a hard ceiling channel that only the operator can set and below that an org default and below that whatever you ask in the message. So for example, watch this channel is capped at three rounds. I'm going to ask them to go back and forth

10 times. We're going to see that since we have that cap on, it's not going to be able to go above three rounds. We can see codeex has started to reply here. And now we have Claude code going for the first time. Okay. And here you can see we have odd code and codeex going back and forth one, two, three different times. And that's when they reached their conclusion. Even though I toolled it to go back and forth 10 times, since

We had the channel capped at three rounds, that is what it did. And if they genuinely finish early, they stop on their own. And there's a pass rule for that. And here is the entire control surface. So this is the channel in the web UI and the ceiling is right there on the conversation. So I could change this from three rounds to two rounds to one round. Uh the org default or custom right here. And that's what the limit for that

Specific channel would be. Now the org default is over here in the admin dashboard. So you can change the rounds and then click apply. And the third layer, you already saw me do this, but this is where we simply just type. for example, go back and forth two times right here. Even though the org default was three, the agents still know to go back and forth twice. And then it has that limit for if you even ask it for 10 back

And forth, their ceiling will win at the channel or the admin level. So, no prompt engineering, no hoping that the model behaves. It's a number that's owned by you. Building this fork surfaced a few real bugs that were already sitting in what they shipped, whether you were using the fork or not. My favorite would be when you revoke a Slack token, which is something that you know an admin may do. The entire platform crash loops every 5 seconds until a human

Fixes it. Not Slack, but the whole platform. The four bugs that we found are shipped in the fork and listed in the guide. So on the left side, we added rooms here. We also added a Slack sync. So you can see the Slack channel right here, as well as the threads. You have your chats here with your projects. And you'll also see we added mentions right here if you want to mention different agents or team members just like you actually would

In Slack or Buzz. This isn't something that the original repo has. You'll also see the highlight of the mentions right here. You can click on the mentions, see that specific agent. Um you'll see the specific agents right here as well. But your Slack channels mirror into the sidebar. So the brain is visible in both places. These are just small things, but they make a difference if you're actually using this tool on the day-to-day. I also made a desktop version as well,

Just like how Buzz has ones. And the guide actually covers both honest ways for you to do. So, the way I'm currently running it right now is a Cloudflare tunnel. So, you can use this if it's just for you, or you could apply it onto something like Fly.io if it is for a team. So, my setup runs both doors at once. My desktop app talks to local host with no sign-in friction and the public host name sits behind Cloudflare access with

Real email signin and that's all in the guide too. So if you want to set it up that way just simply give your AI agent a guide and it will help you do so. Say walk me through the process based on this guide and it will do just that. And in order to set up your Slack agents you're going to want to come to the admin section in onboarding and you're going to want to actually create your QM Slack agent. It's

Going to bring you over here where you can go through the process of actually creating your Slack agent. We'll already have the default manifest set up. You can create it. Once you create it, you can click on enter Slack tokens here. This right here is for your default Slackbot. So, you're going to want to add your bot user off token, your app level token, and then there's also a per agent Slackbot. You can add multiple different agents here. You'll see I

Have my codeex one set up here. And if you need any help with this, just simply just ask your AI agent with the doc. This is also another area that we added to the fork which is where you can see your agents in the web UI here. So we have our QM cloud code right here. Our codecs we can see specific instructions enabled one disabled one. Uh create a new agent. You can put the name here who it's available to um

Different you know settings and then the harness right here. The model by default and the instructions right here. So very similar to Buzz. In the admin section, you can see where we actually assigned our specific Slack bots to the specific agent harness right here. All right, so that is QM and my fork of QM, which raises the obvious question. There's another serious player in this category who shipped something recently. And how do these two tools compare side by side? Now, one

Thing I mentioned about Buzz in my last video is that Buzz has no reliable break. Their own docs admit that the cancel command is broken and rate limiting is a stub. Now QM ships a real break, but it is a dollar budget. It limits what agents spend, not what they do. My fork adds the other one, a turn budget, the operator owns. So to compare the two, Buzz suggests while QM meters and then the fork enforces. So if you watch my

Video on Buzz or you've heard about Buzz, know the team behind it, which is Jack Dorsey. also is calling itself a multiplayer agent harness. And the fascinating part is they looked at the same problem and made opposite bets. Buzz built its own foundation from scratch. A decentralized cryptography signed protocol where every agent and every human gets their own key. It wants to replace Slack and GitHub entirely. Whereas QM did the opposite. It's pragmatic. It runs on a normal database and instead

Of replacing Slack, it moves in and lives inside the Slack you already have. So when I read the buzz source, the thing it truly nailed was identity, proving which agent did what, tying every action back to a human. When I read QM, the thing it nailed was scoping, keeping 20 people's agents from stepping on each other. Different obsessions. And here's what neither of them actually has solved yet. Real orchestration. getting agents to reliably hand work off to each other with hard

Rules about when to stop. Both of them lean on the model to just figure it out. Nobody has enforced it in code yet. That's the next frontier and whoever nails it wins this category. One more thing that made me laugh here is that multiplayer AI was literally in YC's own request for startups this year. They published the brief asking founders to go and build this and they shipped the reference implementation themselves. I saw someone on Hacker News joking that YC is

Now its own portfolio company. So, should you actually go ahead and use QM? Well, let me give you the straight answer plus one detail that tells you everything about where AI tooling is right now. So, straight answer, if you're a technical team that's comfortable self-hosting and you want to give your people the real AI leverage on internal work, UM is one of the most thoughtfully built things in the space right now and it's free open source. Go run it in a

Sandbox, play with it, and it will give you ideas even if you never actually ship it to production. If you're not technical or you're thinking about pointing this at customer data or running it unattended or something that actually matters, I'd probably say not yet. It's version 0.1 and it tells you that itself. So, you got to respect that. And I have to end on this because it is one of the most 2026 AI things I've seen. So, why Cominator open sources

Their own AI agent harness system that runs their entire company. And the most argued topic on the entire hacker news thread is not the architecture, not the security, but a bundled writing skill that bans the mdash, that little long dash, because it's a tell that the text was written by AI. And you can see people argued about punctuation under a launch about the future of work here. Though there is a real critique here buried in the joke that one writing skill

Is around 22,000 tokens of instructions. So that is a large prompt just to sound human. Funny thing is I also do ban M dashes from my agents as well just because of this fact. A quick recap. QM's best idea is scoping and it's genuinely good. Its most honest feature is a security file that tells you exactly what it can't protect and you should actually read it. Its most underrated feature is deploying without forking and it turned this multi- aent category that

Is somewhat new now into a two- horse race kind of between Buzz and QM at the moment. Now, I know a lot of people are building their own internally. We're building one as well at our company, but these are kind of the two big ones that people have open sourced that I've seen so far. So, once again, you can go ahead and clone either their original version or my forked version right here. It's fully open sourced. You can get access to

The entire GitHub repo, the desktop version, as well as this 30-page doc that you can give to your agent to get everything set up in this video in the free school community. Link for that will be in the description down below. I put everything into the guide that you will need so your agent can go ahead set this up. If you have any questions along the way, make sure to ask your agent and we'll help you along this process. But let

Me know your thoughts down below. Have you tried QM yet? How do you feel about it? Do you like it? Do you hate it? How do you think it compares to something like Buzz? And is there anything else you think I missed or any ideas that you want to add? Let me know in the comments down below. And I'm going to be making a lot more videos, guys. So, if you want to stay up to date, make sure to smash that

Like button and subscribe to stay up to date with the future videos. Other than that, guys, that's pretty much it for this video. I will see you in the next video, guys. Keep hustling, keep grinding, and of course, guys, accelerate your stride.