Pick OpenClaw if you want the most chat apps and a project run by an independent non-profit. Pick Hermes Agent if you want an agent that saves its own reusable skills as it works.
Both are free under the MIT licence and run on the same kind of small server, so hosting rarely decides it. Upkeep differs more: from 1 to 27 October 2026, OpenClaw shipped 15 stable releases and Hermes Agent 8.
This comparison draws on both projects' documentation, their security policies and their GitHub pages, read on 27 October 2026.

Who this is for
- you want an always-on AI agent and cannot decide between the two best-known open-source ones
- you care about upkeep and security as much as about features
- you want to know whether the choice changes the server you rent
Not for you if you want a workflow tool with a visual editor: see n8n hosting.
Key numbers
- 29 channels chat apps OpenClaw works with, by its home page Official page,
- 21 surfaces messaging surfaces in Hermes Agent's gateway documentation, the browser included Official page,
- 722 advisories security advisories OpenClaw published on GitHub in 2026, 14 of them critical Official page,
- 15 releases stable OpenClaw releases from 1 to 27 October 2026, against eight for Hermes Agent Official page,
At a glance
Hosting needs are close; upkeep and the security model differ most.
| OpenClaw | Hermes Agent | |
|---|---|---|
| Who makes it | the OpenClaw Foundation, an independent US non-profit | Nous Research |
| Licence and price | MIT; no paid version and no hosted tier | MIT; Nous Portal model credits are optional |
| Written in | TypeScript; needs Node 24.22.1 or 26.12.1 or later | Python; current installs run on Python 3.14 |
| Chat apps | 29, including WhatsApp, Telegram, Slack and iMessage | 21, including Telegram, Discord, Slack, WhatsApp and Teams |
| What stands out | the widest choice of chat apps, and apps for Windows and macOS | saves a reusable skill when it solves a new problem |
| Strangers who message it | most chat apps answer them with a pairing code | turned away unless allowlisted or paired; unknown e-mail senders ignored |
| Security model | one trust boundary per Gateway; sandbox, read-only mode and approvals | the operating system is the only real boundary; two ways to isolate the agent |
| Stable releases, 1 to 27 Oct 2026 | 15, counting its maintenance line | 8 |
| Security advisories on GitHub in 2026 | 722, of which 14 critical and 249 high | none published; private reports accepted |
| GitHub stars on 27 Oct 2026 | about 393,000 | about 258,000 |
| Memory to rent | 1 GB plus swap in its own guide; Kamatera suggests 2 GB | no figure in its documentation |
Hosting: the same small server for either
Both run as a background service on a small server and call an AI model over the internet. OpenClaw's Gateway listens on port 18789 on the server itself. Hermes Agent's gateway is a single background process that its installer sets up as a service.
OpenClaw's own DigitalOcean guide starts at 1 GB of memory with swap, and Kamatera suggests 2 GB for a personal assistant. Hermes Agent's documentation gives no figure, and Contabo sells its plans for both apps from 8 GB.
Hostinger, DigitalOcean, Kamatera and Contabo can install either one when you order a server. Cloudways and Hostinger also run both on managed plans. Our guides to OpenClaw hosting and Hermes Agent hosting compare what each host sets up.
| Plan | Price | New-customer offer | Link |
|---|---|---|---|
| Hostinger KVM 1 1 vCPU, 4 GB | $6.49/month 24-month term, paid up front; Black Friday first-term price, then $11.99 checked 25 Oct 2026 | Black Friday prices on VPS plans and AI automation apps, for the first subscription term only checked 30 Oct 2026 | Visit |
| Hostinger KVM 2 2 vCPU, 8 GB | $8.79/month 24-month term, paid up front; Black Friday first-term price, then $14.99 checked 25 Oct 2026 | Black Friday prices on VPS plans and AI automation apps, for the first subscription term only checked 30 Oct 2026 | Visit |
| DigitalOcean Basic Droplet, 2 vCPUs, 4 GiB 2 vCPU, 4 GB | $24/month billed per second (60-second minimum), capped at this monthly price · VAT not stated checked 25 Oct 2026 | New accounts get a $5 credit that lasts 90 days and covers every product except SaaS add-ons checked 29 Oct 2026 | Visit |
| Kamatera Cloud server Type B (General Purpose), 2 cores, 4 GB 2 vCPU, 4 GB | $39/month monthly, no long-term contract · VAT not stated checked 25 Oct 2026 | A 30-day trial of one cloud server in a configuration worth up to $100 checked 29 Oct 2026 | Visit |
| Contabo Cloud VPS 4 4 vCPU, 8 GB | €4.40/month 24-month term, paid up front; the price applies to the first 24 months and to the EU region in Germany; other regions cost more each month checked 25 Oct 2026 | No offer | Visit |
Some of the lowest prices need a long term paid in advance; the term is under each price.
Updates: how often, and what can go wrong
OpenClaw moves fast. From 1 to 27 October 2026 it shipped 15 stable releases, counting a maintenance line, and the newest, 2026.10.8, arrived on 27 October. Hermes Agent shipped 8, the newest being 0.22.5 on 23 October.
Fast releases bring fixes quickly, but each update is a chance for something to break. An open report in OpenClaw's tracker describes about 8 hours of recovery after an upgrade to 2026.9.5. Two more open reports describe the Gateway's memory growing over days.
OpenClaw's update guide asks for a verified backup before a significant update, and openclaw update --dry-run previews one. Hermes Agent keeps its data, memory and skills in one folder, ~/.hermes, which is what to back up before hermes update.
Security: two different models
OpenClaw's security guide is built on one rule: one Gateway is one trust boundary. The Gateway stays on the server by default, and most chat apps answer strangers with a pairing code. Tools can run in a sandbox or ask for approval.
Its guide also flags two defaults to change. The message tool can send across conversations and chat apps, and the container images listen on an exposed address. Our guide to securing OpenClaw goes step by step.
Hermes Agent's security policy calls it a single-user personal agent. It names the operating system as the only real boundary against a model that misbehaves, since checks inside the agent can be fooled.
For isolation, it offers two set-ups. One runs the agent's shell commands in a container, a remote host or a cloud sandbox; the other wraps the whole agent. Its gateway turns away anyone not allowlisted or paired.
OpenClaw publishes each fixed flaw as a GitHub security advisory: 722 in 2026, 14 of them critical. Hermes Agent takes private reports through the same system but had published none by 27 October 2026.
A long public list means many fixes to apply, and it also shows open reporting. An empty list does not prove a project has no flaws, so keep either agent updated.
Model costs
Neither agent includes a model: you bring an API key, and the same model costs the same with either. OpenClaw also supports signing in with a ChatGPT or Codex subscription. Its documentation still calls an API key the most predictable choice for an always-on server.
Hermes Agent works with your own provider or with credits from Nous Portal, Nous Research's model service. Our guide to what OpenClaw or Hermes Agent costs each month adds the server and the model together.
Choose OpenClaw if...
- you want the widest choice of chat apps, iMessage included
- you want an app for Windows or macOS as well as a server install
- you prefer a project run by a non-profit foundation with no paid tier
- you can keep up with releases several times a week, with a backup before each one
Choose Hermes Agent if...
- you want an agent that saves reusable skills as it solves new problems
- you work in Python, or plan to extend the agent yourself
- you want a security model built on isolating the agent at operating-system level
- you would rather apply fewer releases a month
Where to host either one
Start with a small server from the list above, or a managed plan if you never want to update one yourself. Our guide to where to host an AI agent compares the four routes in general.
How we checked
We read both projects' documentation, home pages and security policies on 27 October 2026. We counted releases, security advisories and GitHub stars through GitHub's own interface on the same day.
Our own measurements will be added here, with their dates, when our test bench runs them.
Mistakes to avoid
Choosing by GitHub stars alone
Instead: Stars measure attention, not fit: compare chat apps, upkeep and security.
Letting strangers message the agent
Instead: Keep pairing or an allowlist on, in either one.
Updating without a backup
Instead: OpenClaw: openclaw backup create. Hermes Agent: back up ~/.hermes.
Reading an empty advisory list as proof of safety
Instead: Keep either agent updated all the same.
Other questions
Is Hermes Agent better than OpenClaw?
Neither is better for everyone. OpenClaw offers more chat apps and a non-profit owner. Hermes Agent saves its own skills and builds its security model on isolating the agent.
Can I run Hermes Agent and OpenClaw on the same server?
Yes: they are separate programs with separate data folders. Give the server enough memory for both, and lock down each one's chat access.
Which one needs less memory?
Neither project gives a firm figure. OpenClaw's DigitalOcean guide starts at 1 GB with swap; Hermes Agent's documentation gives none.
Sources and test notes
- OpenClaw home page, read
- OpenClaw documentation: install, read
- OpenClaw documentation: DigitalOcean, read
- OpenClaw documentation: security, read
- OpenClaw documentation: tool and agent permissions, read
- OpenClaw documentation: updating, read
- OpenClaw documentation: model authentication, read
- OpenClaw issue #153257: update to 2026.9.5 and 8 hours of recovery, read
- OpenClaw releases, read
- OpenClaw security advisories, read
- Hermes Agent home page, read
- Hermes Agent documentation: installation, read
- Hermes Agent documentation: messaging gateway, read
- Hermes Agent security policy, read
- Hermes Agent releases, read
- Kamatera: OpenClaw, read
- Contabo: OpenClaw hosting, read
- Contabo: Hermes AI VPS hosting, read
Every page read on 27 October 2026; releases, advisories and stars counted through GitHub's API the same day.