We set up Hermes Agent in steps: name the job, choose local or hosted, install from the official source, connect one channel and a model provider, keep approvals on, connect accounts narrowly, then test and hand off.

We have set up Hermes Agent for clients, and the process is deliberately unglamorous: agree on one job, decide between a local install and Nous Research's optional hosted cloud agent, install from the official source, connect one messaging channel, connect a model provider, keep dangerous-command approval on, choose an isolated backend for anything unattended, test with real examples, and hand off with notes. Hermes Agent has a lot of features. A good first setup turns on only the ones the job needs. Product facts here come from hermes-agent.nousresearch.com and its official docs, current as of November 2026.

Step 1: name the job

We begin with one job in one sentence, plus what the agent may not do. Hermes Agent can schedule work, delegate to subagents, browse, and run commands. That range is useful later. On day one it is a temptation to do too much. A daily briefing, inbox triage with drafted replies, or a weekly follow-up list are typical first jobs. See What to automate first with AI.

Step 2: local or hosted

Hermes Agent runs on macOS, Windows, or Linux, and Nous Research also offers an optional hosted cloud agent. We help the client choose:

  • Local keeps the agent on a machine the client controls. It needs a computer that stays on for unattended work.
  • Hosted removes the machine to maintain, at the cost of relying on a hosted service.

According to the official FAQ, stopping a local agent ends the process but does not delete saved conversations, memories, or skills; hosted agents are managed separately. We explain that so nobody is surprised later.

Step 3: install from the official source

We use the official desktop app on Mac or Windows, or the official terminal install on Linux. We record the version, where configuration lives, and how to update it. No unofficial builds.

Step 4: connect one channel

Official channels include Telegram, Discord, Slack, WhatsApp, Signal, email, and the command line, plus the desktop app. We pick the one the client already uses all day. Then we set who may talk to the agent. Official security docs describe allowlists and DM pairing rather than open access, and we use them.

Step 5: connect a model provider

The official FAQ says you connect a model provider after install, either your own or Nous Portal for model and tool credits. We help the client pick one they can afford and maintain, and we write down which one is used. We do not resell model credits.

Step 6: keep approvals on and pick a backend

Hermes Agent's official security model is layered: who can talk to the agent, approval for dangerous commands, file-write protections, optional container isolation, and scanning for prompt injection in context files, among others. Our settings for a business:

  • Dangerous-command approval stays on. Official docs warn that modes which disable approval prompts belong only in trusted environments.
  • For a gateway that will run unattended, we prefer an isolated terminal backend. Official docs list options including Docker, SSH, Singularity, and Modal, with container hardening when you use an isolated backend.
  • Secrets stay out of chat and out of files the agent reads. Official docs describe credential filtering in several places; we do not rely on it alone.
  • Third-party skills are added only when needed. Official docs say they are scanned before install. We still review what a skill does.

More in Securing an AI agent: basics.

Step 7: connect accounts narrowly

We connect only what the first job needs, typically one mailbox and one calendar, with the narrowest access that works. The agent can be given more later.

Step 8: instructions, memory, and the approval rule

We write the instructions with the client: what starts the job, where information lives, what good looks like, what never to do, and when to stop and ask. Hermes Agent's persistent memory is useful for preferences and recurring context. We seed it with a few facts the client is comfortable with, rather than everything at once. Our default approval rule is in Approval rules for AI agents.

Step 9: scheduled jobs

If the job runs on a schedule, such as a morning briefing, we set it up using the natural-language scheduling the official site describes, then watch the first several runs. We make sure the client knows how to pause or change the schedule.

Step 10: test and hand off

We test against real examples with approvals on, including awkward cases: an ambiguous request, a message from someone not on the allowlist, and a request that should be refused. Handoff covers how to talk to the agent, approve or deny a risky step, pause it, and where it runs, with written notes the client keeps.

Questions

Is Hermes Agent harder to set up than OpenClaw?

They are different rather than harder. Both are open-source agents you can run yourself. Hermes Agent documents more sandbox backends and an optional hosted agent, which gives more choices to make. See Hermes Agent vs OpenClaw.

Should I use the hosted option?

If you do not want to maintain a machine, the optional hosted cloud agent may suit you. If you want everything on hardware you control, run it locally. We talk through it on the discovery call.

Can Hermes Agent run jobs while I sleep?

Yes, officially it can run scheduled work unattended through its gateway. For unattended work we prefer an isolated backend and approvals on anything risky.

Do you only set up Hermes Agent?

No. We have also set up OpenClaw and Grok Bot for clients.