
Hermes Agent vs OpenClaw: which one should you run?
The main difference between Hermes Agent and OpenClaw is who each one is for. Hermes is Nous Research's agent for one person, built around learning from you and writing its own skills. OpenClaw is a non-profit-run assistant built to reach you across many chat apps and devices, and one install can be shared by a mutually trusting team with the same tool access (not isolated tenants; mutually untrusted users need separate Gateways). Both are free, MIT-licensed and run on your own computer. Both connect to MCP servers, both run cloud or local models, and both install with one command, so the right pick depends on how you'll use it.
What is Hermes Agent?
Hermes Agent is an open-source AI agent made by Nous Research that you install on your own machine or server. Nous describes it as "the self-improving AI agent": it keeps a memory of you, writes its own reusable skills after hard tasks, and can search its own past conversations. You can use it in a terminal, in the Hermes Desktop app, or from chat apps such as Telegram, Discord, Slack, WhatsApp and Signal.
Nous Research is a venture-backed startup founded in 2023. In July 2026 TechCrunch reported it was raising at least $75 million at a $1.5 billion valuation, and that Nous sells a hosted version of Hermes on paid tiers from $20 to $200 a month. The self-hosted agent itself is free.
What is OpenClaw?
OpenClaw is an open-source AI assistant that runs on your own computer and meets you in the chat apps you already use. Peter Steinberger built it with the community, and the OpenClaw Foundation, an independent 501(c)(3) non-profit, now looks after it. The Foundation employs the core team, signs releases and is funded by donors including Amazon, OpenAI and Red Hat. It sells nothing: no paid tier, hosted service or token.
At its centre is the Gateway, a background service that holds your sessions, tools and chat connections. One Gateway can serve one person on a laptop or a mutually trusting team sharing the same delegated tool authority. It is not tenant isolation.
Side-by-side comparison
Side by side, the two share a licence, MCP support and a wide choice of models; they differ most in who makes them and who they're built for. Every row comes from the projects' own docs and GitHub pages, checked on 29 September 2026. The links are under Sources.
| OpenClaw | Hermes Agent | |
|---|---|---|
| Made by | OpenClaw Foundation (non-profit) and community | Nous Research (venture-backed company) |
| Licence | MIT | MIT |
| Written in | TypeScript (Node.js) | Python |
| Built for | One person or a mutually trusting team sharing tool access; not isolated tenants | One person ("a single-tenant personal agent") |
| Chat apps | WhatsApp, Telegram, Slack, Discord, iMessage, Signal, Teams and 20+ more | Telegram, Discord, Slack, WhatsApp, Signal, email, Teams, Matrix, iMessage (via BlueBubbles) and more |
| Apps | Web Control UI, terminal UI, macOS menu bar app, Windows Hub, Linux app, iOS and Android companion apps | Terminal UI, Hermes Desktop (macOS and Windows packages), Android via Termux |
| Memory | Markdown files (USER.md, MEMORY.md, daily notes) with search | MEMORY.md and USER.md with size limits, plus past-session search |
| Learns new skills | Yes, through Skill Workshop (on by default) | Yes, the headline feature, with a curator that tidies old skills |
| Models | Anthropic, OpenAI, OpenRouter, Bedrock and more; local via Ollama, LM Studio, llama.cpp, vLLM | Nous Portal, OpenRouter, Anthropic, OpenAI and more; local via Ollama, LM Studio, llama.cpp, vLLM |
| MCP servers | Yes (stdio, SSE, Streamable HTTP, OAuth) | Yes (stdio and HTTP, plus a Nous-reviewed catalogue) |
| Operating systems | macOS, Linux, Windows (native or WSL2) | macOS (Apple Silicon is top tier), Linux, Windows (native or WSL2) |
| Published minimum | 1 vCPU, 1 GB RAM, about 500 MB disk | None published; the README says a "$5 VPS" works |
| Commands run on | Your computer, unless you turn on sandboxing | Your computer, unless you pick a container or remote backend |
| Releases in September 2026 | Six stable releases (v2026.9.1 to v2026.9.6) by 23 September | Five releases (v2026.9.7 to v2026.9.24) |
| GitHub stars (29 Sept 2026) | 390,817 | 250,132 |
How you talk to each one
Both agents talk to you through chat apps, so day to day they feel much the same. OpenClaw leans harder on devices. It has companion apps for iPhone, Android, Mac, Windows and Linux that add voice, camera and screen access, and its FAQ describes a common setup: the Gateway on an always-on Mac mini, with your laptop paired as a "node". It also supports a shared Gateway for mutually trusting teammates with the same tool access, with sessions that show who started and steered them; this is not tenant isolation.
Hermes offers a full terminal interface, a desktop app and a single gateway process for your chat apps. Its docs suggest typing /new at natural breaks in a chat so memory works well, because a Telegram chat otherwise stays one long session for weeks.
Memory, skills and learning
Learning is Hermes's headline feature, but OpenClaw now does it too. Hermes keeps two small memory files that load at the start of every session, and it saves hard-won workflows as skills with its skill_manage tool. A background "curator" marks unused skills stale after 30 days and archives them after 90, and never deletes them outright.
OpenClaw also keeps memory in plain Markdown files, including a USER.md profile and dated daily notes that the agent can search. Its self-learning feature, Skill Workshop, reviews longer conversations and turns fixes and successful work into skills. By default it applies them automatically. You can switch it to "propose" to approve each one, or turn it off.
Both follow the AgentSkills format for skills. The Hermes Skills Hub can even install skills from ClawHub, OpenClaw's skill registry.
Models: cloud and local
Both work with nearly any model provider, including local ones. Hermes calls Nous Portal, its own subscription, "the recommended way to run Hermes Agent": one login covers 300+ models plus web search, image generation and voice. You don't need it, though. Hermes works with any OpenAI-compatible endpoint, including Ollama running on your own machine.
Both projects warn about small local models. Hermes needs at least 64,000 tokens of context for agent work. OpenClaw's FAQ says a local model is "usually no" good for casual chat, because small models drop context and raise the risk of prompt injection. OpenClaw's managed llama.cpp setup checks your hardware first, and its smallest recipe needs 8 GiB of memory.
MCP and tools
Both connect to MCP servers, subject to each client's supported transport, authentication and tool requirements. Protocol support alone does not guarantee that every server or tool works with either agent. OpenClaw lets you add servers from its settings page, its command line or its config file. Hermes reads them from config.yaml and ships a catalogue of MCP servers that Nous staff have reviewed.
Installing Hermes Agent and OpenClaw
Each installs with one command. For Hermes on Linux, macOS or WSL2, run curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash. On Windows, run iex (irm https://hermes-agent.nousresearch.com/install.ps1) in PowerShell. Then run hermes setup, or just hermes to start chatting. Mac and Windows users can instead download the Hermes Desktop app, which bundles the agent and the software it needs.
For OpenClaw on macOS, Linux or WSL2, run curl -fsSL https://openclaw.ai/install.sh | bash. On Windows, run iwr -useb https://openclaw.ai/install.ps1 | iex. The installer adds Node.js if you don't have it, then opens a setup wizard that checks your model access and configures the Gateway. There's also a macOS app and a Windows Hub app.
Hardware
Neither needs a powerful computer unless you run models locally. OpenClaw publishes a minimum of 1 vCPU, 1 GB RAM and about 500 MB of disk, recommends 2 GB or more, and has a Raspberry Pi guide (Pi 4 or 5, 64-bit OS). Its FAQ says plainly that you don't need to buy a Mac mini. You only need a Mac for Mac-only features like iMessage. See whether a Mac mini is worth it for OpenClaw.
I couldn't find a published RAM minimum for Hermes; its README says it runs on "a $5 VPS". Its top-tier platforms are Apple Silicon Macs, Windows 10 and 11, and 64-bit Linux on x86 or ARM. If you're unsure whether a computer you already own will do, try Which computer should run your agent? or the short answers for Hermes on a Raspberry Pi, Hermes on a VPS, OpenClaw on a Raspberry Pi and OpenClaw on a VPS.

Cloud-model agent, always on, cheapest box with 16 GB that doesn't feel slow. Intel Core i5-9500T (6-core), 16 GB memory, 256 GB NVMe SSD.

Always-on cloud-model agent on macOS (needed for OpenClaw's iMessage), silent and 4 W at idle. Apple M6, 12-core CPU, 12-core GPU, 153 GB/s, 16 GB memory, 256 GB SSD.
Amazon links on this site are affiliate links: if you buy, I may earn a commission, at no cost to you. As an Amazon Associate I earn from qualifying purchases. Nobody pays to be picked.
Security defaults
Out of the box, both run the agent's commands directly on your computer with your permissions, so treat both with care. Both block strangers by default: an unknown person who messages the bot gets a pairing code you have to approve.
Hermes asks an extra AI model to judge risky shell commands before they run. Clearly dangerous commands are refused and unclear ones come to you for approval. Its security policy is blunt, though: "The only security boundary against an adversarial LLM is the operating system." Nous recommends running the whole agent in a container or NVIDIA OpenShell if it reads the open web, email or untrusted MCP servers.
OpenClaw's own docs say sandboxing and command approvals are off by default. It offers Docker, Podman, SSH and OpenShell sandboxes, plus an openclaw security audit command. It supports "one trust boundary per gateway", so people who don't trust each other need separate Gateways. There's more in Is OpenClaw safe?
Updates and stability
In the September 2026 snapshot used for this comparison, both projects were shipping updates about weekly. OpenClaw offered stable, beta and dev channels, plus a slower "extended-stable" channel that never installed updates by itself. That OpenClaw channel information has not been rechecked for this October update.
Hermes update, checked 19 October 2026: the v0.22.0 notes, published 14 October, say source installs now follow tested SemVer stable releases by default, rather than every commit on main. The notes retain an opt-in main channel and say installer reruns preserve the channel you chose. Check your installation's selected channel before updating.
21 October patches: check your update route
Hermes published two stable patches on 21 October. The v0.22.3 notes describe a roll-up since v0.22.2 and defer the full curated notes to v0.23.0. The v0.22.4 notes describe removing desktop suggestion pills, making the approval-mode chip follow the active profile and fixing WhatsApp re-pairing after logout. These are maintainer-reported changes, not fixes tested on this site's machines.
The update route matters. Both releases say the CLI can use hermes update or the official installer, while the Microsoft Store app updates through the Store. They explicitly say the desktop app installed from an .exe or .dmg installer is unchanged in these patches and updates with the next bundled release. A CLI patch announcement therefore does not establish that an installed desktop bundle has changed. Check the version actually running before attributing a result to the patch.
The approval-mode chip is a display of the active profile's mode, not proof that a particular action was approved. These release notes do not establish a fix for every gateway timeout, provider error or session-history issue. Keep the reboot example below separate from the newer patch changes.
Hermes after a reboot: check the service, then the task
Release update, 19 October 2026. Hermes Agent's v0.22.2 release notes, published 15 October, describe a gateway startup fix. After a Linux reboot, an unrelated process could receive an early process ID that Hermes mistook for its own running gateway. A supervised gateway could then refuse to start. The maintainers say Hermes now records the boot it started in and, when that record cannot settle the question, checks a host lock held by a live gateway.
That is a targeted startup correction. It does not prove that your unattended tasks will recover after a restart. Before depending on a spare computer or server, use this check on a noncritical installation:
- Record what you run. Note the installed version, installation method and service manager. Back up configuration through your installation's documented procedure, keeping credentials private.
- Read the upgrade notes first. The v0.22.2 release warns that sessions which ran under the 0.22.1 gateway cannot be deleted in 0.22.2. Preserve session history; do not clear it to make a test pass. Follow the current supported update route for your installation and confirm the version actually running.
- Restart during a maintenance window. Check the service manager and logs for a healthy gateway. Then send a harmless request through the chat channel you normally use and verify its reply there.
- Check scheduled work separately. If you rely on a schedule, test one harmless scheduled task after reboot. Record a startup failure, missing response or repeated work separately; a healthy service alone does not establish a completed task.
This is a proposed acceptance check, not a result from my machines. I have not reproduced this startup bug or tested the fix locally. The comparison table above remains a 29 September snapshot; this section updates only the dated release information.
The same release describes changes for long local-model turns on laptops, including prompt-reading progress and handling time spent asleep. Test your own model server and task before drawing a reliability conclusion. These notes do not establish that you need a faster computer. If you want help getting an agent running and checking its behavior, see the agent setup packages.
Which should you pick?
Pick by who will use the agent: OpenClaw for a trusted person or team sharing the same access across many devices, not tenant isolation; Hermes for one person who wants the agent to learn. In more detail, pick OpenClaw if:
- more than one mutually trusting person will use the agent with the same tool access, such as a family or a small team. This is not tenant isolation. Mutually untrusted users need separate Gateways, ideally with separate OS users or hosts. Shared Gateways are a supported setup, while Hermes's security policy calls it a single-tenant personal agent.
- you want companion apps on your phone and your other computers that feed voice, camera and screen to one agent.
- you want a small, documented starting point like a Raspberry Pi or a 1–2 GB VPS.
- it matters to you that a non-profit with no paid product runs the project.
Pick Hermes Agent if:
- it's just you, and you want an agent whose main design goal is learning your habits and building its own skills.
- you'd rather pay one subscription (Nous Portal) than collect separate keys for models, search, images and voice.
- you want a company with investment behind it and a paid hosted option if self-hosting gets tiring.
My take: the gap is narrower than most comparison posts suggest, because the two projects have grown towards each other. Both can even import the other's memory and skills. Choose on who will use it and how much tinkering you want, then keep it on a stable channel. To budget for either, see what OpenClaw really costs to run. The model bill works the same way for Hermes.
Neither quite right? Four more OpenClaw alternatives, from a 10 MB Go program to an agent that runs every task in its own container, with the computer each one needs.
Questions
Is Hermes Agent better than OpenClaw?
Neither is better for everyone. Hermes is designed for one person and puts learning first. OpenClaw can be shared across many chat apps and devices by mutually trusting people with the same tool access, not isolated tenants. Mutually untrusted users need separate Gateways. As of September 2026 both learn skills, support MCP and run local models.
Can I switch from OpenClaw to Hermes, or back?
Yes. hermes claw migrate imports OpenClaw's persona, memories, skills, messaging settings and allowlisted API keys. OpenClaw's openclaw migrate hermes imports Hermes's config, memories, skills and supported credentials, and shows you a preview first.
Do Hermes Agent and OpenClaw cost money?
Both are free, MIT-licensed software. You pay for whatever AI model you connect, unless you run a local one. Nous Research also sells Nous Portal and a hosted Hermes, and TechCrunch reported paid tiers from $20 to $200 a month. The OpenClaw Foundation has no paid tier.
Hermes vs OpenClaw hosting: where should each one run?
Either one runs on a computer you already own or on a small rented server. OpenClaw publishes a minimum of 1 vCPU, 1 GB of RAM and about 500 MB of disk, and has a Raspberry Pi guide. Hermes publishes no minimum; its README says a "$5 VPS" works. If you don't want to run a server, Hostinger sold managed hosting for either one at $5.99 to $10.99 a month in September 2026, and TechCrunch reported that Nous sells a hosted Hermes on paid tiers. The OpenClaw Foundation has no hosted plan. See what it costs to run.
Does Hermes Agent run on Windows?
Yes. Native Windows 10 and 11 is a top-tier (Tier 1) platform for Hermes, installed with a PowerShell command or the desktop app. WSL2 also works. OpenClaw also runs on native Windows or WSL2.
Do I need a Mac mini for either?
No. Both run on Linux and Windows as well as macOS. OpenClaw's FAQ says a small VPS, home server or Raspberry Pi-class box works. You only need a Mac for Mac-only features such as iMessage.
Which one is more popular?
On 29 September 2026, OpenClaw had 390,817 GitHub stars and 82,203 forks. Hermes Agent had 250,132 stars and 53,401 forks. Both projects run Discord servers for help.
Sources
- GitHub: openclaw/openclaw README, licence and star count (checked 29 September 2026)
- GitHub: NousResearch/hermes-agent README, licence and star count (checked 29 September 2026)
- OpenClaw docs: What is OpenClaw?
- OpenClaw docs: Why OpenClaw (governance and default security)
- OpenClaw docs: Install
- OpenClaw docs: Platforms
- OpenClaw docs: FAQ on providers, hardware and hosting
- OpenClaw docs: Raspberry Pi
- OpenClaw docs: Memory overview
- OpenClaw docs: Self-learning
- OpenClaw docs: Connect MCP servers
- OpenClaw docs: Local models
- OpenClaw docs: Security trust model
- OpenClaw docs: Release channels
- OpenClaw docs: Migrating from Hermes
- GitHub: OpenClaw releases (checked 29 September 2026)
- OpenClaw Foundation
- Hermes Agent docs: Installation
- Hermes Agent docs: Platform support
- Hermes Agent docs: Updating and uninstalling
- Hermes Agent docs: Persistent memory
- Hermes Agent docs: Skills system
- Hermes Agent docs: MCP
- Hermes Agent docs: LLM and model providers
- Hermes Agent docs: Messaging gateway
- Hermes Agent docs: Security
- GitHub: Hermes Agent security policy
- GitHub: Hermes Agent releases (checked 29 September 2026)
- TechCrunch: Hermes agent maker Nous Research in talks for new funding at $1.5B valuation (13 July 2026)
How to use this guide
The recommendations in this guide draw on the linked documentation, stated requirements and dated price examples. They are not a scored performance benchmark or a promise that a particular setup will work for every task. Check the source date beside a claim before buying or changing a system.
Start with the task, the software it needs and a computer you already own. Compare the whole cost: hardware, model usage and time to check the work. Hardware purchases may earn this site an affiliate commission; paid setup is optional.
Updated : clarified the comparison method and reading paths. Older price snapshots and software-source checks have not been refreshed by this edit.