Hermes Agent Explained: Nous Research's Self-Improving AI
Hermes Agent is an open-source, self-improving AI assistant from Nous Research. See how its memory, skills, and gateway work for small business teams.
OpenClaw is a free, open-source AI assistant that runs on your own computer and meets you in chat apps. Learn what it does, what it costs, and setup reality.
OpenClaw is a free, open-source AI assistant that runs on your own computer and talks to you inside the chat apps you already use. In January 2026 it went from 9,000 to more than 60,000 GitHub stars in 72 hours, one of the fastest runs in open-source history, and it now sits near 389,000 stars. This guide explains what it is, what it does for a business, and what setup honestly takes.
If you want the other side of the open-source story first, read our Hermes Agent explainer, or jump to whether open-source agents are worth the setup for a small team.
OpenClaw began as a personal project by developer Peter Steinberger and the community around it. It was first called Clawdbot, then renamed Moltbot, and finally OpenClaw, which is the name it carries today. The renames can be confusing when you find older tutorials online: a guide written for Moltbot in early 2026 generally still applies, but check that the commands match the current release, v2026.9.4 from September 2026.
The project is now stewarded by the OpenClaw Foundation, an independent US nonprofit, and the code is published under the MIT license: you can run it for business and modify it without paying anyone. There is no paid tier and no official hosted service. One fact worth knowing because it gets misreported: donors include the University of Michigan, OpenAI, Amazon, and Red Hat, but OpenAI is a donor, not an owner.
The star count deserves honest interpretation. Roughly 389,000 GitHub stars by September 2026 makes OpenClaw enormously popular by open-source standards. Stars measure developer interest, not business readiness: treat the number as evidence of a large community writing skills and guides.
Strip away the excitement and OpenClaw is straightforward: an assistant that lives on your machine, remembers context, and acts on your instructions in plain language. Its persistent memory keeps track of your preferences and past work across sessions. Its skills library, more than 100 preconfigured AgentSkills, bundles reusable procedures so you do not start from zero on common tasks. A "skill" here just means a saved way of doing something the agent can load and run.
Concretely, OpenClaw can browse the web, fill out forms, read and write files on the connected machine, and run shell commands, which are text instructions to the computer that can do anything from organizing files to installing software. It runs cron and background tasks, meaning scheduled jobs that execute on a timetable and report back, like a morning briefing or a periodic check of something you monitor. One Gateway, the single program that holds all the connections, can serve an individual owner or a whole shared team.
For a service business, the realistic jobs are internal: drafting inbox replies for approval, assembling a morning schedule brief, monitoring supplier pages or review sites, preparing quote follow-up lists, and keeping notes organized. It is an operations helper, not a phone system. If the problem you feel most is missed calls and slow booking, that is customer-facing work for a receptionist or voice agent, not for a self-hosted assistant.
OpenClaw's most practical advantage is reach. It supports 29 channels in total, including WhatsApp, Telegram, Discord, Slack, Signal, iMessage, and Microsoft Teams, with native apps for macOS, iOS, Android, Windows, and Linux. You talk to the same assistant from whichever app is closest.
For a small team, this removes the usual adoption fight. Nobody needs to learn a new dashboard. The owner checks in from a chat app on the phone, the office manager uses the desktop app or Slack, and a technician in the field sends a text-style message asking what is next on the schedule. Because one Gateway serves the whole setup, the assistant has the same memory and skills no matter which channel a question arrives through.
There is a trade-off hiding in this convenience. Every connected channel is another door into a system that can read files and run commands, so connect only the channels your team uses, with proper authentication on each. Our self-hosted security checklist covers this in detail: fewer doors, each with a lock.
The price structure has three parts, and only the first is zero. The software is free. The AI model it connects to bills by usage, so your monthly cost scales with how much the assistant works. Hosting is the third part: you need a computer that stays on, whether that is a machine in the office, a server you rent, or a packaged deploy like DigitalOcean's one-click hardened setup starting at $24 per month. "Hardened" means preconfigured with security basics, which is worth preferring over a bare manual install.
Setup takes a technical person somewhere from an afternoon to a couple of days: installing the Gateway, connecting a model provider with billing and usage alerts, linking your channels, loading skills, and configuring the security posture described below. Ongoing costs are attention rather than money. Someone must apply updates, review what the assistant remembers, and revisit permissions when staff change. If that sounds like more operations than you want, say so early: managed automation exists precisely so you can skip it, and our open-source agents for small business guide walks through the comparison honestly.
OpenClaw can run shell commands with full access or inside a sandbox, and this single setting deserves your attention. Full access means the assistant can do anything your user account can do, which is convenient and genuinely risky: a mistaken instruction or a malicious message that tricks the agent could reach further than you intended. Sandboxed execution confines it to a contained area with limited reach, which is safer and occasionally means a task needs your help to finish. For business use, start sandboxed and widen access only for specific, understood needs.
Layer the standard controls on top. Give the assistant its own dedicated credentials, never your personal logins, and grant each connection the least privilege it needs. Put approval gates on anything irreversible: payments, messages to customers, deletions, and credential changes should wait for a human yes. Keep logs of what it did and update promptly when releases land.
OpenClaw fits owners who have, or can hire, a little technical support and who value control: your data stays on your machine, your model choice is yours, and no vendor can change the terms underneath you. Marketing agencies, consultancies, trades with an ops-minded office manager, and teams that enjoy tinkering all do well with it. If you already run Hermes Agent, the two overlap heavily, and our Hermes comparison helps you decide whether to switch or stay.
Skip it for now if you need customer-facing results this week, if nobody will own updates and access review, or if your actual problem is one defined workflow like answering calls or chasing quotes. Those are solved faster by purpose-built automation. Open-source assistants reward owners who want the engine; everyone else should buy the finished vehicle.
What is OpenClaw?
OpenClaw is a free, open-source personal AI assistant that runs on your own computer and meets you inside chat apps you already use, including WhatsApp, Telegram, Discord, Slack, Signal, iMessage, and Teams. It is stewarded by the OpenClaw Foundation, an independent US nonprofit, and published under the MIT license with no paid tier and no hosted service.
Is OpenClaw really free?
The software is free, but running it is not. You pay for the AI model it connects to through usage-based API billing, plus the computer or server it runs on. DigitalOcean offers a one-click hardened deploy starting at $24 per month. Budget for model usage, hosting, and setup help if nobody on your team runs servers.
Is OpenClaw safe for a business to use?
It can be, with deliberate setup. OpenClaw can run shell commands with full access or inside a sandbox, and that choice decides your risk. Use sandboxed execution, dedicated credentials with least privilege, approval gates on sensitive actions, and regular updates. Our self-hosted security checklist walks through each control before you connect business data.
What is the difference between OpenClaw and Hermes Agent?
Both are free, MIT-licensed assistants that live in chat apps, but OpenClaw is known for its 29 channels, large community skill library, and viral open-source momentum, while Hermes Agent emphasizes a self-improving learning loop of memory and autonomous skills. Hermes even ships a migrate command for OpenClaw setups. See our Hermes Agent vs OpenClaw comparison for the full breakdown.
List the two or three recurring internal tasks you would hand to an assistant first, then decide whether you want to own the setup or buy it done. Our free six-step AI automation plan turns that list into priorities: start your AI automation plan. If you want a second opinion on self-hosted versus managed for your team, book a call.
Hermes Agent is an open-source, self-improving AI assistant from Nous Research. See how its memory, skills, and gateway work for small business teams.
Hermes Agent and OpenClaw are both free, open-source AI assistants. Compare their channels, features, costs, and migration path to pick the right fit.
Seven business jobs OpenClaw handles well, from inbox triage to team briefings. Each includes a setup sketch, time savings, and the guardrails to run it safely.
More articles: browse the full Praktivo blog.