Abacus Hermes and OpenClaw Setup: Two Pre-Baked Agents, One Job Each

The Abacus Hermes and OpenClaw setup is two pre-baked agents on SuperComputer, and the vendor sentence is "or," not a fused pair. Claw is Abacus's hosted OpenClaw. Hermes is the Nous Research agent. I give one of them a job, or I leave both off.

I'm William Spurlock, founder of Spurlock Studios LLC, an AI Systems Architect and Fractional AI CTO. I've built 600+ automations with 500+ still live, spent 20,000+ hours on agentic systems, and watched clients drop 35,000+ hours of busywork. I still do not staff a Monday as one chat that "does AI."

The September seat post already split research writing, overnight ops, and the repo. This page is the always-on slice that post left alone. The May showdown already compared self-hosted Hermes, OpenClaw, and Agent Zero. If you want that choice, read Hermes vs OpenClaw vs Agent Zero. If you want which Abacus seat I actually occupy on a writing day, read when I use Abacus instead of Grok Bot or Cursor. This page is the click inside ChatLLM Teams, and the credit meter under it.

I read the Abacus help pages, the Hermes docs, and the OpenClaw docs on October 5, 2026. Prices and meters below are those pages, not a quote I invented.

---

What is the Abacus Hermes and OpenClaw setup? #

The Abacus Hermes and OpenClaw setup is Hermes plus Claw, two pre-baked apps on Abacus AI SuperComputer, with Claw documented as the hosted version of OpenClaw. The Abacus AI Agent sandbox does not get that row. Pro is the door.

The SuperComputer guide, as I read it on October 5, 2026, has a section titled "Pre-Baked Apps: Hermes and Claw." Both are "ready to deploy in a click" on that VM. The same page's comparison table marks "Persistent AI agents (Hermes, OpenClaw)" as available on SuperComputer and unavailable on the Abacus AI Agent. You open the computer from the dashboard or at supercomputer.abacus.ai.

The Abacus Claw how-to, read the same day, uses a plainer sentence. Abacus Claw is "the hosted, managed version of OpenClaw," and you can start at claw.abacus.ai with no local setup. So the name "OpenClaw" on the comparison table and the name "Claw" on the pre-baked list are the same product family. They are not a third agent.

Upstream, the projects are not Abacus inventions.

• Hermes Agent docs, read October 5, 2026, call it the self-improving agent built by Nous Research. It creates skills from experience and keeps memory across sessions.
• OpenClaw docs, read the same day, call it a self-hosted gateway from the OpenClaw Foundation, MIT licensed. You run the Gateway. Config lives at ~/.openclaw/openclaw.json.

Abacus is selling a hosted seat for each of those, on a computer they operate. That is a different operational fact from cloning the repo yourself.

Name on the bill	What Abacus says it is	Where I open it
Hermes	Pre-baked agent on SuperComputer, with memory, skills, and continuous research	supercomputer.abacus.ai
Claw	Hosted, managed OpenClaw	claw.abacus.ai, and also listed as a SuperComputer pre-baked app
Abacus AI Agent	Task sandbox for apps, docs, and scheduled tasks	ChatLLM Agent mode. The comparison table says persistent Hermes and OpenClaw are not this seat

Three surfaces get mashed in one login, and they are not the same job:

1. The model picker answers a prompt and stops.
2. The Abacus AI Agent finishes a task in a sandbox and hands you an artifact.
3. SuperComputer keeps a machine on, and the pre-baked agents are the always-on option on that machine.

If you only needed a draft, you never had to open door three. The seat post is still my rule for drafts. This post starts when the artifact has to exist after you close the laptop.

---

Why is Hermes and Claw easy to miss inside ChatLLM? #

Hermes and Claw are easy to miss because they sit behind Pro, on SuperComputer, while the page people buy is a model picker. Basic does not include the computer. There is no secret toggle on the $10 plan.

The billing FAQ, read October 5, 2026, splits the subscription like this.

Plan	Price on that page	Credits on that page	SuperComputer, Hermes, Claw
Basic	$10 per user per month	20K	Not included. Agent access is limited to 3 tasks
Pro	$20 total per month, Basic plus the extra $10	30K, which includes the extra 10K from Pro	Included. The FAQ names persistent agents such as Hermes and Claw

The Claw how-to repeats the Pro bundle in its own billing table: $20 per user per month, 30,000 credits, Abacus AI Agent plus Abacus Claw. I treat those two pages as the same shelf, not as two prices.

People miss the pair for a boring reason. ChatLLM is sold as "stop paying five model tabs." The always-on agents are a second product with a runtime meter. They look like a bonus bullet. They behave like a server.

The path I use when someone asks me to turn this on:

1. Open billing and confirm the plan says Pro, not Basic.
2. Open supercomputer.abacus.ai, or claw.abacus.ai if the job is the channel agent.
3. Pick Hermes or Claw. Do not arm both for the same digest.
4. Leave the Always On toggle off until the credit log for one real day is in front of me.

The SuperComputer guide says the VM auto-shuts down when nothing is running, and that an Always On toggle in settings keeps it up. I want that toggle off on minute one. A bonus bullet should not become a month of runtime before the job is even named.

---

Which job should Hermes own, and which job should Claw own? #

Hermes should own work that ought to become a skill. Claw should own work that ought to show up in a chat channel on a schedule. I do not point both at the same inbox.

That split follows the pages, not a vibe. On the SuperComputer guide, Hermes "excels at research, planning, and workflows that require persistence and learning." The bullets under it are persistent memory, reusable skills generated from experience, and continuous operation for research, monitoring, or long analysis. Claw, on the same page, is "a centralized agent that connects to over 50 messaging platforms, including WhatsApp, Slack, and Telegram," with flows and tools and a chat interface.

The Claw how-to is stricter about day-one setup than the "over 50" sentence. Its channel table names three: WhatsApp, Telegram, and Slack. Its tool examples name Slack, Notion, GitHub, Gmail, Google Calendar, and Google Drive. I configure the three named channels first. I do not tell a client the other forty-seven are done because a sentence said "over 50."

Upstream Hermes is wider than the Abacus blurb. The skills page, read October 5, 2026, says the agent creates, updates, and deletes skills with the skill_manage tool. A skill is a SKILL.md procedure, loaded on demand. The memory page says memory lives in ~/.hermes/memories/ as MEMORY.md and USER.md, injected at session start. Skills hold the procedures, and memory holds what the agent knows about you. The docs homepage also lists a messaging gateway, Telegram, Discord, Slack, WhatsApp, Teams, and "20+ platforms." The Abacus Hermes blurb does not repeat that gateway list. I do not assume the pre-baked button installed every platform until I see the channel in the UI.

OpenClaw's own homepage, same day, lists Discord, Google Chat, iMessage, Matrix, Microsoft Teams, Signal, Slack, Telegram, WhatsApp, Zalo, and more, through channel plugins. Abacus Claw's setup section still walks WhatsApp, Telegram, and Slack. Abacus hosts that copy, and its setup page documents three channels. I follow the documented path.

Job in front of me	Agent I arm	Why
Telegram or Slack digest on a clock	Claw	The how-to is channels, cron, and tools
A weekly source check that should be reused next month	Hermes	The product description is skills plus memory
A git diff I will ship	Neither. Cursor	Same rule as the seat post
Overnight ops that must run after the laptop closes	Neither, on my Monday. Grok Bot already has that seat	I am not moving that seat in this post

A Monday I will not run: Claw summarizes Gmail at 8am, and Hermes also summarizes Gmail at 8am. One job, two memories, two chances to disagree, one credit meter still burning.

A Monday I will run if the ask is specifically this surface:

• Claw owns one channel and one cron. Example from their own page shape: a morning digest to Telegram.
• Hermes owns one procedure that should land as a skill, such as a weekly source check with the steps written down.
• Cursor owns the repo if a file in git has to change.
• Grok Bot keeps overnight ops. I said that in September. I am not quietly reversing it because Abacus added a button.

Claw has one behavior I want in the assignment, because it surprises people. The how-to says past sessions are read-only. You cannot keep typing in an old session. You start a new one, and persistent memory is what carries forward. The OpenClaw Dashboard on that same page auto-refreshes every 5 seconds so you can see live sessions, cron status, and connected tools. The dashboard is the check. The old chat tab is not the memory.

The self-host install for Hermes, if you are not using the Abacus click, is the Hermes field guide. This page does not repeat that install.

---

What does the Abacus Hermes and OpenClaw setup cost while the computer stays on? #

Pro is $20 per user per month with 30,000 credits, and the computer then costs 1 credit per 5 minutes while it runs. Model calls are a second bill. The $20 line is the subscription. It is not a promise that always-on is free inside the credit pool.

Two official pages describe the stop button differently. I am not going to sand them into one sentence.

• The SuperComputer guide says you pay for active runtime, the VM auto-shuts down when nothing is running, and Always On is how you override that.
• The Claw how-to says the agent's cloud computer costs 1 credit per 5 minutes "regardless of activity," and "if you want to stop charges, you must explicitly stop the computer."

If you are on claw.abacus.ai, follow the Claw sentence and stop the computer. If you are on SuperComputer, read the Always On toggle before you leave the tab. I do not know from these pages whether Claw's "dedicated cloud computer" and the SuperComputer VM are one meter or two. Budget one meter per computer you leave running, and read the credit log on day one.

The rate itself is the same number on both pages: 1 credit per 5 minutes. Here is the arithmetic, labeled as arithmetic, not as an Abacus quote.

• One hour is 60 divided by 5, which is 12 credits.
• One day is 24 times 12, which is 288 credits.
• Thirty days of continuous runtime is 30 times 288, which is 8,640 credits.
• A 30,000 credit Pro pool minus 8,640 leaves 21,360 credits before any model call.

Computer left running	Credits at 1 per 5 minutes	Left of 30,000 before model calls
1 hour	12	29,988
24 hours	288	29,712
30 days continuous	8,640	21,360

That remainder is not "free chat." The SuperComputer guide says two more bills sit beside hosting.

1. Usage of the Abacus AI Agent or the CLI inside the VM is billed separately from hosting.
2. The VM carries an ABACUS_API_KEY for RouteLLM. LLM calls with that key consume subscription credits separately from the hosting charge.

The Claw how-to adds a third escape that is only half an escape. Bring your own API key and model calls bill to that key. Hosting still costs 1 credit per 5 minutes. A personal OpenAI or Anthropic key does not turn the computer off.

The billing FAQ also splits credit types. Purchased credits roll over. The monthly 20K that comes with the subscription resets to 20K on the renewal date. I am not going to claim the extra Pro 10K follows the rollover sentence. The FAQ names the 20K allotment in that sentence, and I will leave the bonus where the sentence left it.

What I want on a scrap of paper before Always On:

• The job, in one line.
• The agent, Hermes or Claw, not both.
• The channel or the skill, one of them.
• The stop rule. Auto-shutdown, or an explicit stop, matching the page you actually opened.
• The day-one credit log, hosting line and model line, before anyone calls it cheap.

If the log surprises you, the setup is wrong. Raising the credit pack is not the fix.

---

When should I self-host Hermes or OpenClaw instead of Abacus? #

Self-host when you need the config file and the box on hardware you control. Use the Abacus click when you want their computer and you accept the credit meter. I do not use the hosted button as a way to skip reading what the upstream project is.

Need	Abacus click	Self-host
Start without a local install	Claw how-to says claw.abacus.ai, no local setup	You run the installer and the Gateway or the Hermes process
Config on your disk	Their UI and their VM	OpenClaw config at `~/.openclaw/openclaw.json`. Hermes memory at `~/.hermes/memories/`
Who bills the runtime	1 credit per 5 minutes while their computer runs, plus model calls	Your VPS bill and your model bill. Abacus is not in that invoice
Channel list I will promise	The three the Claw how-to names, until I see more in the UI	Whatever that project's docs list and I have actually connected

I skip Abacus for this job when any of these are true:

1. The operator needs ~/.openclaw/openclaw.json on a machine they SSH into.
2. Skill writes have to be approved on hardware they control, which is the Hermes write-approval gate in the Nous docs, not a toggle I found on the Abacus blurb.
3. The credit log after one weekend is already the story, and the job has not shipped.
4. The artifact is a git repo. That stays on Cursor.

OpenClaw's docs are explicit that the point of the project is self-hosting. You run a Gateway on your machine or a server. Abacus Claw is a hosted copy with their models pre-wired and their security wording on the how-to. Hosted does not mean "the upstream project stopped being a gateway you were supposed to own." It means you traded the box for a credit meter.

Hermes is the same trade. The Nous install is a script on Linux, macOS, WSL2, or Windows. The Abacus button puts that agent on their VM. If the thing you wanted was the learning loop, the button can be the right door. If the thing you wanted was the files under ~/.hermes/, the button is the wrong door.

I still send the framework choice, self-host versus self-host versus Agent Zero, back to the 2026 decision post. This page only answers the hosted door inside one vendor.

---

FAQ #

What is the Abacus Hermes and OpenClaw setup? #

It is two pre-baked agents on Abacus AI SuperComputer: Hermes, and Claw, which Abacus documents as hosted OpenClaw. The SuperComputer guide, read October 5, 2026, lists both under pre-baked apps, and its comparison table says persistent Hermes and OpenClaw are not available on the Abacus AI Agent sandbox. Pro is required.

Why is Hermes and Claw easy to miss inside ChatLLM? #

They are easy to miss because SuperComputer is a Pro feature, and Basic never shows that door. The billing FAQ, read October 5, 2026, prices Basic at $10 per month with 20K credits and no SuperComputer. The model picker is the product people see first. The always-on agents are a runtime meter behind the upgrade.

Which job should Hermes own, and which job should Claw own? #

Hermes owns procedures that should become skills. Claw owns chat channels and cron. The SuperComputer guide describes Hermes with memory, reusable skills, and long research, and describes Claw as the messaging agent. I do not run both against the same Gmail digest.

What does the Abacus Hermes and OpenClaw setup cost while the computer stays on? #

Pro is $20 per user per month with 30,000 credits, then 1 credit per 5 minutes while the computer runs, with model calls billed separately. Thirty continuous days at that rate is 8,640 credits of hosting, which leaves 21,360 of the 30,000 pool before a model call. That arithmetic is mine, from the rate on the SuperComputer guide and the Claw how-to, both read October 5, 2026.

When should I self-host Hermes or OpenClaw instead of Abacus? #

Self-host when you need the config on your disk and a machine you control. OpenClaw's docs, read October 5, 2026, describe a self-hosted Gateway with config at ~/.openclaw/openclaw.json. Hermes memory lives under ~/.hermes/memories/. Use Abacus when the click and their credit meter are the trade you actually want.

Does ChatLLM Basic include SuperComputer, Hermes, or Claw? #

No. Basic does not include SuperComputer, and the pre-baked agents live on SuperComputer. The billing FAQ, read October 5, 2026, lists Basic at $10 per month with 20K credits and Agent access limited to 3 tasks. Pro is $20 total per month with 30K credits, and that is the tier that names Hermes and Claw.

Is Abacus Claw the same product as OpenClaw? #

Abacus Claw is Abacus's hosted, managed version of OpenClaw, not the self-hosted Gateway itself. The Claw how-to, read October 5, 2026, uses that sentence and points at claw.abacus.ai. Upstream OpenClaw, on docs.openclaw.ai the same day, is software you run, MIT licensed, from the OpenClaw Foundation.

Can Hermes and Claw share one SuperComputer? #

The SuperComputer guide lists both as pre-baked apps on that computer, and I still would not arm both for one job. I have not seen a page that prices them as two separate credit meters versus one shared VM. Budget one meter per computer you leave running, then check the credit log on day one before you assume they share a bill.

How many credits does thirty days of always-on runtime cost before model calls? #

At 1 credit per 5 minutes, thirty continuous days is 8,640 credits of hosting. That is 12 credits an hour, times 24, times 30. Subtract from the 30,000 Pro credits and 21,360 remain before model calls, Agent usage, or CLI usage inside the VM. Those other lines are extra, per the SuperComputer guide read October 5, 2026.

Does a bring-your-own API key stop the Abacus hosting charge? #

No. The Claw how-to says your key takes the model calls, and hosting still costs 1 credit per 5 minutes. I read that page on October 5, 2026. Stopping the charge means stopping the computer, or relying on SuperComputer's auto-shutdown if that is the surface you opened and Always On is off.

Which chat channels does the Abacus Claw guide name for setup? #

The Claw how-to's channel table names WhatsApp, Telegram, and Slack. The SuperComputer blurb says Claw connects to over 50 messaging platforms, including those three. I promise the three the setup page walks through. I do not promise the rest until they show up in the UI I am looking at.

Can the Abacus AI Agent sandbox run persistent Hermes or OpenClaw? #

No. The SuperComputer comparison table marks persistent AI agents (Hermes, OpenClaw) as a SuperComputer capability, not an Agent capability. The Agent is the task sandbox. The computer is the always-on machine. I read that table on October 5, 2026.

Do past Abacus Claw sessions stay open for more chat? #

No. The Claw how-to says past sessions are read-only, and new work starts a new session. Persistent memory is what carries forward. The OpenClaw Dashboard on that page auto-refreshes every 5 seconds if you want live session and cron status instead of scrolling an old tab.

Should overnight ops move off Grok Bot onto Abacus Claw? #

Not on my Monday. Grok Bot keeps overnight ops, and Claw only gets a channel job I have named. I wrote that seat split in the September Abacus post. A new pre-baked button does not move a seat I already assigned.

---

I am William Spurlock. I build these systems at Spurlock Studios LLC. On an AI automation strategy call, bring the job in one line and, if you already turned the computer on, the day-one credit log. I will tell you whether Claw, Hermes, or neither should be armed, and whether the stop rule matches the page you opened. I will not invent a savings number to make the button feel cheap.