What Is a QM Agent? (And How I'd Actually Run One)

Do not index

This image illustrates the architecture of a YC multiplayer quartermaster (QM) system designed to facilitate company work. Central to the system is a server cluster managing shared memory, file crates, and a durable Postgres state database. Multiple workspaces including a Jane bot desk, legal project room, Slack channel wall, marketing room, and web portal screen are connected to this central hub. These components interact via shared memory and file crates, enabling real-time collaboration across different company functions. The image emphasizes the integration of people, rooms, Slack, and web interfaces, highlighting the QM's role in organizing and managing resources effectively. The service is marketed with managed QM available from $3.99 per month, branded as Agent 37.

A QM agent is Y Combinator's open-source multiplayer harness for company work in Slack and on the web. If your local pilot already works and you want managed hosting without assembling Postgres and Slack yourself, Agent 37 is currently the only managed QM host. Browser plans start at $3.99/mo. Self-hosting stays free (MIT) plus cloud and ops time.

Friday was my last day thinking QM was "just another agent repo."

It is not.

YC open-sourced it in July 2026 under MIT. Official page: qm.ycombinator.com. Code: github.com/yc-software/qm. Quartermaster is the naming joke that stuck. The metaphor is load-bearing. QM is for company-shaped fleets, not one personal butler.

I hosted QM first on Agent 37 for a reason. We already ran OpenClaw, Hermes, Claude Code, Codex, OpenCode, and Grok CLI on the same isolation model. When YC shipped QM, the gap was obvious: nobody was offering managed hosting for it. So we did.

What is a QM agent?

A QM agent is a quartermaster-style multiplayer harness: people and project rooms get isolated memory, files, permissions, and sandboxes, with Slack and web as native surfaces. YC open-sourced it in July 2026 under MIT and is honest that it is early. It is for company-shaped fleets, not one personal butler.

YC's own path matters.

They tried a simple internal agent loop. OpenClaw raised expectations for personal agents. They ran 50+ Hermes agents as employee assistants and felt fleet admin pain. QM is what they built next: Hermes-like flexibility with a stronger company admin story.

What ships in practice:

Multiplayer by design: people and rooms, not one shared personality

Slack and web first, not afterthought plugins

Crons, webhooks, shared memory and files, browser support

Harness-agnostic core with adapters for Pi, OpenCode, Codex, and Claude Code

Postgres for durable state; TypeScript / Node core

If you only need one WhatsApp butler or one coding CLI, QM is usually the wrong first buy.

If you need agents for people and projects inside how the company already works, it is the interesting new option.

How is a QM agent different from OpenClaw or Claude Code?

This image visually compares three different software tools: QM, OpenClaw, and Claude Code, highlighting their distinct deployment units and use cases. QM is depicted as a personal multi-channel assistant interacting through apps like WhatsApp, Telegram, and Discord on multiple personal devices. OpenClaw is shown as a repository coding command-line interface (CLI), represented by multiple monitors displaying code and a central server stack. Claude Code is illustrated as a company multiplayer environment, with multiple teams working in an office setup connected via Slack, supported by drones for communication or coordination. The image emphasizes these tools serve different jobs and are not substitutes, guiding users to select based on their preferred deployment unit. The bottom section includes the tagline 'Pick by unit of deployment' and credits 'Agent 37' as the creator.

OpenClaw is a personal multi-channel agent computer. Claude Code is a repo coding CLI. QM is a company multiplayer harness (Slack + web) that can drive coding backends. Different jobs. Do not treat them as ranked substitutes. Pick by the unit of deployment you actually need.

Question	OpenClaw	Claude Code	QM
Unit of deployment	Usually one personal stack	One coding workspace	Many agents across people and projects
Primary interface	Broad chat apps	Terminal / IDE	Slack + web
Admin story	You operate the gateway	Session must stay open for channels	Fleet admin is part of the thesis
Model posture	BYOK	Anthropic-native	Harness-agnostic backends

I do not rank these. I run all three for different jobs.

OpenClaw docs sit around personal residency. Claude Code docs sit around Anthropic repo work (Anthropic Claude Code). QM sits where "Jane's bot" and "the legal project room" both need isolation and an admin story.

The four differences that matter:

Unit of deployment. OpenClaw: one personal stack. Claude Code: one coding workspace. QM: many agents across people and projects.

Primary interface. OpenClaw: chat apps. Claude Code: terminal / IDE. QM: Slack + web first.

Admin story. Hermes and OpenClaw can scale. YC built QM because fleet management hurt at 50+ agents.

Model lock-in. Claude Code is Anthropic-native. OpenClaw is BYOK. QM's core is harness-agnostic.

How do I run a QM agent in production?

Two honest paths: self-host into your VPC, or use managed QM so you are not on-call for the quartermaster. Production means always-on, durable Postgres, secret hygiene, webhook reachability, and isolation per person or room. Local success does not equal those five.

I've run production agents three ways: laptop, DIY VPS, managed host.

The laptop died every time the lid closed.

The VPS cost about $6 to $12/mo plus the Sunday I spent chasing Slack webhooks.

Managed is what I ship when uptime is the product.

Should I self-host a QM agent?

Self-host if you need the bits in your cloud, you have platform capacity, and owning the fork is a real requirement. Clone the repo, read SECURITY.md before you romanticize Slack bots with tool access, and budget for Docker-shaped infra, Slack Bolt, and early open-source sharp edges.

Always-on checklist:

A host that does not sleep when a laptop lid closes

Durable Postgres

Secrets for models and Slack

Network path for webhooks and Slack events

Isolation so agent A cannot paw through project B's files

An operator who will read logs when a cron misfires at 2 a.m.

Gotchas I would budget for on day one:

YC says it is early. Treat production as a pilot with sharp knives.

Security is not optional. Execute sandboxes plus company connectors are not a chat toy threat model.

Fleet complexity arrives fast. Ten agents feels cute. Fifty is a product.

Swappable backends are a feature and a footgun. Great until nobody remembers which room is on Codex vs Claude Code when the bill spikes.

What does managed QM on Agent 37 cost?

Agent 37 is currently the only managed QM host. You get an always-on isolated instance on the same substrate used for OpenClaw, Hermes agent, Claude Code, Codex, OpenCode, and Grok CLI. BYOK for models. One personal Free agent sleeps when idle; paid browser plans start at $3.99/mo.

When we launched OpenClaw hosting at $0.99 per user per month, we got 30+ paying customers in a day. Timing mattered. Price mattered. We had roughly 200 competitors doing OpenClaw hosting. We won because we were early and damn cheap.

QM is the same playbook with a scarcer supply: today we are the only managed host.

Live managed plans from Agent 37 pricing:

Plan	Price	Specs
Free	$0/mo	1 vCPU, 4 GB RAM, 2 GB; sleeps when idle
Basic	$3.99/mo	1 vCPU, 4 GB RAM, 8 GB
Plus	$9.99/mo	1 vCPU, 4 GB RAM, 12 GB
Pro	$29.99/mo	2 vCPU, 8 GB RAM, 20 GB, Chromium
Max	$99.99/mo	4 vCPU, 16 GB RAM, 40 GB

Cloud API is white-label by default and meters per minute. An always-on floor (2 vCPU, 4 GB RAM, 4 GB disk) is about $4.76/mo awake 24/7. Idle compute drops to disk only (about $0.36/mo for 4 GB at $0.09/GB).

This image presents a detailed comparison of Managed QM hosting plans available from Agent 37, starting at $3.99 per month. The visualization features four server towers representing different hosting tiers: Basic ($3.99/mo) with 1 vCPU, 4 GB RAM, and 8 GB storage; Plus ($9.99/mo) with 1 vCPU, 4 GB RAM, and 12 GB storage; Pro ($29.99/mo) with 2 vCPU, 8 GB RAM, and 20 GB storage with Chromium support; and Max ($99.99/mo) with 4 vCPU, 16 GB RAM, and 40 GB storage. Additional graphics indicate connectivity and integration features such as Slack and web APIs, highlighting the cloud API availability at approximately $4.76 per month. The image is designed to clearly communicate hosting service options and technical specifications for prospective customers or knowledge base documentation.

At $50/hour, one afternoon of DIY Slack + Postgres babysitting exceeds a year of Basic ($47.88) before tokens.

Agent 37 powers agents for 100+ companies. That is the credibility check I care about more than a landing-page adjective.

Product page: QM agent hosting on Agent 37. Same dashboard: agent37.com.

Which QM agent setup fits which use case?

This image illustrates five distinct scenarios where different QM agents from the Agent 37 platform are recommended based on the task environment. It depicts workspaces ranging from slack web company fleets to personal multi-channel setups, repo coding, lean overnight operations, and managed always-on server environments. Each scenario is paired with a specific agent: QM for company fleets, OpenClaw for personal multi-channel, Claude Code for repo coding, Hermes for lean overnight ops, and a managed option for always-on setups starting at $3.99/month. The visual emphasizes matching the appropriate QM agent harness to the workload, highlighting the principle of one job per trust boundary.

Pick by job, not by hype. Many teams run more than one harness on isolated instances. Confused bots that try to be Slack ops, WhatsApp butler, and senior engineer invent a fourth job: incident response for the agent. Match the harness to the work, then pick a host.

Job	Pick
Company Slack/web multiplayer agents	QM
Personal multi-channel assistant	OpenClaw
Shipping code in a repo	Claude Code (or Codex / OpenCode)
Lean overnight ops / memory	Hermes
Always-on host without DIY ops	Managed instance on Agent 37

I shut down 13 startups. The ones that died slow usually tried to make one product do three jobs.

Same pattern with agents.

Sister reads: OpenClaw vs Claude Code and Claude Code hosting.

What breaks first when a QM agent leaves the laptop?

The first breaks are usually uptime, secrets, Slack event delivery, and cross-room isolation, not model quality. Your local pilot already proved the prompt loop. Production fails on the boring infrastructure that kept working while your laptop stayed open and reachable.

Concrete failure modes I watch for:

Sleeping host. Laptop lid, idle Free-tier sleep, or a VPS without restart policy. Slack thinks the bot is ghosting.

Postgres without backups. Shared memory and room files look fine until the disk dies.

Webhook path. Slack Bolt and cron triggers need a stable URL and TLS story.

Secret sprawl. Model keys and Slack tokens in screenshots, shared .env dumps, or one key across fifty rooms.

Isolation shortcuts. One shared volume "for convenience" is how legal's draft lands in marketing's agent.

Managed hosting does not erase those risks. It means someone else owns the box staying awake, the runtime updates, and the isolation substrate. You still own BYOK hygiene and what each room is allowed to do.

Frequently Asked Questions

Is QM open source?

Yes. MIT. Code at github.com/yc-software/qm. YC frames it as early and experimental.

QM agent vs Claude Code?

Different jobs. Claude Code is a coding CLI. QM is a multiplayer work harness that can drive coding backends including Claude Code.

How much does a QM agent cost to run?

Software is free. You pay tokens plus a host. Managed on Agent 37 starts at $3.99/mo plus BYOK tokens (Free sleeps when idle). Tokens often dominate once the fleet gets chatty.

Who made QM?

Y Combinator. Official page: qm.ycombinator.com. Quartermaster is the naming joke that stuck.

Is Agent 37 the only managed QM host?

Today, yes for managed hosting. Self-hosting into your own cloud remains available from the open-source repo.

Where should I start this week?

Read the official page and SECURITY.md. Decide fleet harness vs personal/coding runtime. Pilot one person or one project room before you invite fifty people.

Start here: QM agent hosting on Agent 37 from $3.99/mo.

What Is a QM Agent? (And How Do I Run One in Production?)