QM vs OpenClaw vs Hermes: Which Agent Fits One Person or a Team

OpenClaw and Hermes are built for one person. YC's QM gives each teammate and Slack channel its own memory and computer. How to choose, and what each costs.

•

Published on

•

QM vs OpenClaw vs Hermes: Which Agent Fits One Person or a Team
Do not index
QM, OpenClaw, and Hermes Agent are all open-source AI agents under the MIT license, and all three can live in Slack. The difference is who they're built for. Pick OpenClaw or Hermes when one person wants a personal assistant. Pick QM, Y Combinator's agent harness, when a team shares the agent: it gives each person and each Slack channel its own memory, files, and computer.
This guide compares the three on who they serve, where they run code, what they take to self-host, and what they cost. We run all three on Agent37, so the hosting numbers below come from our own plans.
We also tested QM on October 26, 2026. A new hosted QM was running 1.4 seconds after we asked for it, answered its first message in 4 seconds, and wrote and ran a Python script on its own computer in 6 seconds.

QM vs OpenClaw vs Hermes at a glance

ã…¤
QM
OpenClaw
Hermes Agent
Made by
Y Combinator
OpenClaw Foundation
Nous Research
Built for
A team
One person, with a team mode
One person
Where you use it
Slack and a web app
WhatsApp, Telegram, Slack, Discord, iMessage, and 20+ more
Telegram, Discord, Slack, WhatsApp, Signal, and a terminal
Memory
Separate for each person and each room
One per agent
One per agent, built around its owner
Where commands run
A durable sandbox per person and per room
On the host, unless you turn on sandboxing
Local, Docker, SSH, Modal, Daytona, or Vercel Sandbox
To self-host
Fly.io, AWS, or one Docker host, plus Postgres
Any Mac, Linux, or Windows machine
Any Mac, Linux, or Windows machine
GitHub stars
15.5k
393k
257k
Star counts are from GitHub on October 26, 2026. QM is the youngest by far: its first version shipped July 29, 2026, and version 0.1.15 on October 20.

The real difference: one person or a team

OpenClaw and Hermes start from one owner. Hermes says it "builds a deepening model of who you are across sessions," and every chat app you connect reaches that same agent. OpenClaw is the same by default: one Gateway, one assistant, your credentials.
OpenClaw does offer a team deployment, and its docs are direct about the tradeoff: "Everyone who can message a tool-enabled agent shares that agent's delegated tool authority." Teammates get their own logins, roles, and private skills, but separate memory means separate agents, and for people who don't trust each other the docs say to run one gateway per tenant.
QM starts from the team. In its own words, "each person and each room has its own scoped memory, files, keychain view, permissions, crons, web apps, and durable sandbox." Your private agent knows your work, the #launch channel's agent knows the launch, and neither reads the other's files unless someone shares them. An admin picks the models, the security posture, and the skills the whole company gets.
That's why the choice is usually easy. If you're the only person talking to the agent, QM's scopes add setup without adding much. If five people will talk to it in Slack, QM's scopes are the main reason to use it.

When to pick QM

Pick QM when:
  • Several people will use the agent in Slack channels, group messages, and projects.
  • Each person's memory, files, and logged-in accounts should stay theirs.
  • You want one admin panel for models, security posture, and shared skills.
QM runs every turn through a central core backed by Postgres. Its agent uses Pi or Claude Code as the harness, and Pi runs models from many providers, so you can switch models without changing anything else. Every command runs in the scope's own sandbox, where installed tools stay installed.
The tradeoffs: QM only works in Slack and on the web, so there's no WhatsApp or Telegram. It's three months old. And self-hosting it is real infrastructure work, covered below.

When to pick OpenClaw

Pick OpenClaw when one person wants an assistant in every app they already use. It reaches more than 25 chat apps, has native apps for macOS, iOS, Android, Windows, and Linux, and has the largest community of the three.
The tradeoff is safety by default. OpenClaw's README says tools "run on the host for the main session unless you configure sandboxing," so an agent that reads a malicious message can act on the machine it runs on. Turn sandboxing on before you connect channels that strangers can message.

When to pick Hermes

Pick Hermes when you want an agent that gets better at your work over time. It creates skills after complex tasks, improves them as it uses them, and searches its own past conversations. It has a built-in scheduler for unattended jobs, and six terminal backends, so its commands can run in Docker or a remote sandbox, away from your laptop.
Like OpenClaw, Hermes is built around one owner. A team can share a Hermes agent in Slack, but it has one memory and one view of who it's working for.

What each one costs to run

Self-hosted. OpenClaw and Hermes run on any small machine. Hermes says a $5 VPS is enough, and your model usage is the main cost.
QM needs more. On Fly.io, QM's own templates give the core a 2 vCPU, 4 GB performance machine, which Fly lists at $66 a month, plus Managed Postgres, which starts at $38. That's at least $104 a month before the web app, portal, sign-in service, storage, model usage, and agent computers. On AWS it's Terraform, an RDS database, and Lambda MicroVMs, priced by your account.
Hosted on Agent37. All three run on the same plans. QM's README and qm.ycombinator.com link to Agent37 as a third-party hosted version of QM.
Plan
Price
What you get
Free
$0
One agent that sleeps after 15 idle minutes, $0.20 of the default model to start
Basic
$3.99/month
Stays on, bring your own model keys, $1 a month in AI credits
Plus
$9.99/month
Free use of GPT-6 Luna, DeepSeek V4.1 Flash, and Mercury 2
Pro
$29.99/month
2 vCPU, 8 GB RAM, built-in browser, $5 a month in AI credits
Max
$99.99/month
4 vCPU, 16 GB RAM, team management, $20 a month in AI credits
QM's computers bill separately, by the minute, from your Agent37 Cloud balance. At QM's default size of 2 vCPU, 4 GB of RAM, and 8 GB of disk, each costs $0.0070 an hour while it works and $0.72 a month while it sleeps. Adding a card gives you $5 of credit and charges nothing.

Try QM on Agent37

  1. Go to agent37.com/qm and sign in. Verify your email if asked.
  1. Click Start QM free.
  1. Click Open QM to chat, or Admin panel to connect Slack and pick models.
  1. Add a card when you want QM to run code on its own computers.
Here's what our test run measured, using the same API the page calls:
Step
Time
QM created and running
1.4 seconds
Web app answering
15 seconds
Signed in as admin
19 seconds
First reply
4.0 seconds
First command, including creating its computer
7.2 seconds
Write and run a Python script
6.0 seconds
The computer QM created for us had 2 vCPU, 4 GB of RAM, and Python 3.11. The script printed the first 10 primes, and QM replied with exactly that output.
Already running QM yourself? You can keep your deployment and move only the computers to Agent37. Set the sandbox backend to agent37 in your deployment config and add your API key as a deployment secret:
{
  "sandbox": { "backend": "agent37" }
}
AGENT37_API_KEY=sk_live_...
QM creates one Agent37 computer per scope, puts it to sleep when it's idle, and deletes it when the scope is torn down. To change the size, set AGENT37_CPUS, AGENT37_MEMORY_GB, and AGENT37_DISK_GB in the core's environment.

Problems to avoid

  • Check the security posture. Hosted QM starts in Dangerous, so Slack and scheduled work never stall waiting for approval. To have a person approve tool calls, switch to Strict or Auto under Admin > Governance. QM's command policy blocks things like recursive deletes in every posture.
  • Connected apps belong to everyone. On hosted QM, an account one person connects, such as Gmail, can be used by every member and every scheduled job. Connect shared accounts, and keep personal ones out.
  • Free QM sleeps. It wakes when you open it on the web, which is fine for trying it. Slack and scheduled jobs need it awake, so move to a paid plan before your team relies on it.
  • Teammates need a way to sign in. Slack works once it's connected. Web sign-in for other people needs an email sender, such as a Resend key.
  • Sandbox OpenClaw and Hermes before strangers can reach them. Both can run commands on the machine they're installed on.

FAQ

Is QM better than OpenClaw?

For a team, usually yes, because each person and channel gets its own memory and computer. For one person who wants an assistant in WhatsApp or iMessage, OpenClaw fits better.

Can a team share OpenClaw or Hermes?

Yes, but everyone shares the agent's memory and tool access. OpenClaw's docs recommend separate gateways for people who don't trust each other.

Who makes QM?

Y Combinator. It's published as yc-software/qm on GitHub under the MIT license, and the first version came out on July 29, 2026.

Is QM free?

The code is free. Running it costs hosting and model usage. Agent37's free plan runs one QM at no charge, with $0.20 of model usage to start.

Which models do they support?

All three work with models from many providers. QM uses Pi, which runs models from many providers, or Claude Code. OpenClaw supports hosted and local providers, and Hermes supports Nous Portal, OpenRouter, OpenAI, and your own endpoint.

Can I run more than one?

Yes. On Agent37, each agent takes its own plan seat, and the free plan covers one agent at a time.

Start with the one that fits

If one person needs an assistant, start OpenClaw or Hermes free. If your team needs one in Slack, start QM free. It takes about a minute.
For a walkthrough of QM's chat and admin screens, read What is a QM agent. For a deeper look at the other two, read Hermes vs OpenClaw vs Claude Code.
Vishnu

Written by

Vishnu

Founder at Agent37, which runs managed hosting for OpenClaw and Hermes agents for 1,000+ users. Writes about what actually breaks when you leave an AI agent running.