Claude Code vs OpenClaw or Hermes: what does an always-on agent add?

Comparing Claude Code with OpenClaw or Hermes starts with what Anthropic already runs for you. Its cloud already runs scheduled Claude Code sessions on your GitHub repos, and Cowork runs scheduled tasks over your connected apps, both with your laptop shut. Neither runs a job on a schedule more often than once an hour, runs a script without calling a model, or answers you on WhatsApp. Hermes Agent and OpenClaw can do all of that from a server that stays on while your laptop sleeps. We also paused and stopped Hermes in the middle of its schedule on October 20, to see which runs a sleeping machine misses.

What Anthropic runs with your laptop closed

Running it in Anthropic's cloud

A cloud session keeps running after you close your laptop. It counts toward your plan's usage limits. Anthropic doesn't charge separately for the cloud VM. Cloud scheduled tasks, which Anthropic used to call routines, start those sessions on a schedule, an API call or a GitHub event. A schedule can't run more often than once an hour, and the feature is in research preview.

When your computer will be off, Anthropic's mobile page points you to cloud sessions or a project. On Pro and Max, an @Claude mention in Slack starts a cloud session.

Cowork's scheduled tasks run remotely too, even with your computer asleep or the Claude Desktop app closed. They run hourly, daily, weekly or on weekdays, using your connectors and the files saved to your Claude account. A task that needs local files or apps can't run anywhere but your computer. On Pro and Max, tasks you made before October 6 that run on your computer keep running there.

On your own computer

Schedules on your own computer won't run while it's asleep. Desktop's scheduled tasks run only while the app is open and the computer is awake. Its Keep computer awake setting stops idle sleep, but closing the lid still puts a laptop to sleep. On wake, Desktop starts exactly one catch-up run for the most recent missed time. So Anthropic warns that a task scheduled for 9am might run at 11pm if your computer was asleep all day.

A /loop task fires only while its session is open and expires after 7 days. Remote Control and Dispatch drive your own machine, so it has to stay on.

What an agent on a server adds

Hermes and OpenClaw keep running on a server you rent, so they're up when your laptop isn't. They'll work with your files there and a model you pick.

Hermes runs one gateway process that connects every chat app you set up, keeps the sessions and runs the cron jobs. Its memory files load at the start of every session, so what it learned in a Slack chat is there when you message it from WhatsApp. A plain cron job on any server runs scripts on time too, and that's fine if all you need is the script. Hermes adds the chat side, since a no-agent job's output arrives word for word in the chat you picked. OpenClaw's list of chat apps overlaps with Hermes's and adds a few, such as Zalo. If you're choosing between the two agents, our guide to which of OpenClaw or Hermes Agent to run covers the rest of that choice.

What a sleeping machine does to a schedule

On October 20 we gave Hermes Agent two jobs on a test VM (not a DreamHost server), one every 5 minutes and one every 15. Both were scripts that write the time to a log, so there wasn't a model involved or a key set. We paused it for 25 minutes to stand in for a closed laptop lid and let it run again. Then we stopped its container for 20 minutes to stand in for a machine that's switched off, and started it again.

About four five-minute runs fell inside the pause and about three inside the stop. Hermes didn't replay any of them. Each job ran once on waking, and Hermes counted the next run from that catch-up. Its job list tagged the 15-minute job's late runs catch-up after missed fire. That matches Hermes's cron docs, which say a long outage collapses into a single run. OpenClaw's docs describe a similar rule that we didn't test: a gateway that wakes after a deadline merges the missed ticks, and on startup it reschedules overdue agent jobs instead of replaying them. If a late run would do more harm than a skipped one, such as a reminder that lands hours after it mattered, hermes config set cron.catch_up_missed false turns catch-up off.

So on a laptop, Hermes wouldn't do any better than Desktop: a 9am briefing on a laptop that sleeps until noon runs once at noon, while the same job on a server that stays up runs at 9.

Running Claude Code and an always-on agent together

The agent can hand coding work to Claude Code when they're on the same server. Hermes ships a Claude Code skill that delegates coding tasks to the claude command. Its preferred path is print mode (claude -p). OpenClaw can reuse your Claude Code login through the claude program installed on the same host. Its openclaw attach command starts Claude Code tied to one OpenClaw session, so Claude Code can use OpenClaw's tools.

1. Start the agent on its server

   Our Hermes Agent and OpenClaw pages say the agent arrives running. From there, our Hermes Agent page has you add an API key from your model provider, and our OpenClaw page has you open the dashboard through an SSH tunnel. Then you connect your chat apps.

2. Install Claude Code and sign in as yourself

   Anthropic's terms don't prevent you from signing in to the unmodified Claude Code program with your own subscription, including where a platform hosts it. With no browser on the server, claude setup-token makes a one-year token.

3. Hand coding jobs to claude -p

   Turn on Hermes's Claude Code skill, or point OpenClaw at the installed claude program. Print mode can't stop to ask you. A tool call that needs approval is denied unless something approves it, so list the tools it may use with --allowedTools, as the skill's examples do.

Stack 4's 4 GB has room for both. On a 4 GB test VM on October 16, Hermes idled at 557 MiB and OpenClaw at 1,024 MiB. One Claude Code run waiting on a reply used about 210 MiB. Adding those separate readings gives about 767 MiB for Hermes with one Claude Code run, or about 1,234 MiB with OpenClaw. We didn't run the agent and Claude Code together. If you'd rather run a team of Claude Code agents than one assistant, our guide to Claude Code agent teams and when to add Paperclip covers that route.

What the model costs on top of the server

The server's price doesn't change with how much the agent works. Your provider bills the model, which our Hermes Agent page calls two separate bills. On our October 16 test install, each Hermes model call from the command line carried a fixed prompt of about 12,200 tokens. So a chatty agent's calls add up. Our guide to what a month of Hermes Agent costs on each model plan prices one session a day on Claude Sonnet 5.5 at about $7.82 a month through OpenRouter.

Whether your Claude plan can pay for the agent's model calls depends on the plan and the route.

Anthropic's support article, updated October 7, says You can still use the Claude Agent SDK, claude -p, and third-party apps with your subscription limits.

The same update says Max and Team plans now include monthly API credits. They cover the Claude API, and claude -p only when it runs with an API key from the Console organization linked to your plan. Anthropic's credits article puts them at $100 a month on Max 5x and $200 on Max 20x, for building your own apps and agents. Everyone with an API key in the linked Console organization draws on that balance, and Pro isn't eligible.

Anthropic's legal page says third-party developers may not offer Claude.ai login in their own apps or route requests through Free, Pro or Max credentials on behalf of their users. Anthropic has updated its support article twice since June, so read it again before you connect a plan.

When to add Paperclip on top

Paperclip sits above the agents and hands them work. It works with Claude Code, OpenClaw, Hermes and others, and wakes them for assigned work or on a schedule. Paperclip gives each agent a monthly budget, with a warning at 80% and a hard pause at 100%, as our Paperclip page describes. The same page says Claude Code, Codex, OpenCode and Gemini come pre-installed in the container. On the October 16 test VM, Paperclip idled with a Hermes worker at about 1,465 MiB and with an OpenClaw worker at about 2,175 MiB. Stack 4 holds either pair at idle, so you can try the pair there. Our guide to whether you need Paperclip on top of OpenClaw or Hermes picks Stack 8 once the agents start working, and covers the wiring.

How we measured

DreamHost ran the schedule test on October 20, 2026. The VM ran under Colima on an Apple-silicon Mac, with an arm64 processor, 2 vCPU and 8 GB of memory. We didn't use a DreamHost server. Hermes Agent v0.22.2 ran from Nous Research's official image with no model key and no provider. We created both jobs with hermes cron create --no-agent --script and read the times from their log and from hermes cron list. A docker pause stood in for the lid. We used docker stop and docker start for switching the machine off and on.

Pausing a container freezes every process but leaves the clock and network running, which a real laptop sleep doesn't, so read the result as Hermes's catch-up rule at work. It's one shortened run of one agent. Our every-5-minutes job ran about every 6 minutes throughout this run. We didn't test OpenClaw's schedule, a run over several days, Claude Code started from the agent, or Anthropic's own schedulers, which this guide describes from their docs. Anthropic's scheduling, cloud, mobile, channels and plan pages were read again on October 21, after the routines rename. The memory figures and the 12,200-token prompt come from our October 16 runs on a 4 GB VM of the same kind. We counted the prompt as the system prompt plus the definitions of Hermes's 21 default tools.

Claude Code and always-on agent FAQs