Hermes Agent vs Grok Bot

Hermes Agent and Grok Bot are the same idea (named persistent bots that message each other) sitting on opposite operating models. Hermes Agent is Nous Research’s open-source roster you host on your own machine, while Grok Bot is SpaceXAI’s hosted cloud-computer teammate that keeps working when the laptop is closed.

I am Adam Stewart. I run Digital Bond, and I write for founder-led Australian businesses that want a system running, not another prompt to babysit.

Published 28 August 2026. Product access, pricing and sources were checked on 27 August 2026.

As of that research cutoff, both products give you a named roster. The useful question is not which brand is smarter. It is isolation and control versus a hosted computer and less ops.

What is Hermes Agent, in plain English?

Hermes Agent is the open-source agent from Nous Research. People searching Hermes AI usually mean this project. You install it. You run it. The software is MIT-licensed (Copyright 2025 Nous Research, confirmed on the Hermes Agent GitHub). The models, the hardware, and the time it takes to keep it healthy are yours.

Most people meet Hermes Agent as a desktop app on macOS, Windows, or Linux. The same runtime also talks through Telegram, Discord, Slack, WhatsApp, Signal, email, and a CLI. You are operating an agent on a machine you control, not renting a chat window.

Bot Mode is what turned that agent into a roster. It shipped as a bundled, default-on desktop plugin in Hermes Agent v0.20.3 (tag v2026.8.16.2). The release is dated 16 August 2026 and was published on GitHub on 17 August 2026 by teknium1. That patch also bundled the teammate protocol, MCP 2.x, Cua Driver 0.20 computer use, and cron self-heal. Full curated notes are promised for v0.21.0.

A Bot in Hermes is not a nickname on a shared account. A Bot is a Hermes profile at ~/.hermes/profiles/<name>/, with its own config, memory, skills, credential configuration, and chat history. Official Bot Mode docs still default new Bots to a shared OAuth/token pool with the main profile, so the folder is not a hard login boundary.

Each Bot can have its own role, model, memory, skills, and avatar. Different Hermes agents on the same roster can pin different models and providers side by side. You write a custom SOUL.md for the persona you actually want.

The canonical Bot Chat is a forever-chat. Type /new and Hermes reroutes it to /compact. Fresh working context. Same relationship.

Routines are Hermes cron jobs, namespaced [bot:name], and they show up in hermes cron list. Group chats seat two to six Bots, with a cap of three serial rounds and ten messages per send. @mentions pull a specialist in. @user escalates a judgement call to you (needs-you badge). Bots also message each other with the message_agent tool. The protocol lives in agent.bot_mode_protocol in config.yaml, and it is on by default.

Cross-machine work uses Desktop Connections as a relay, or hermes peer for always-on gateway-to-gateway messaging with no desktop in the loop.

You operate the runtime. The default install is local. Official docs also ship an optional web dashboard with password or OIDC login (Okta and others) and a dashboard login audit file, not a hosted enterprise console.

On 17 August 2026, @NousResearch posted the official Bot Mode line:

Introducing Bot Mode for Hermes Desktop.

Your agent profiles become a series of named Bots. Each Bot has its own role, model, memory, skills and profile picture; Bots can use any model and even communicate with each other.

Build a specialist Bot once to use it forever.

Nous Research (@NousResearch), 17 August 2026 https://x.com/NousResearch/status/2089429432612147572

On that post, as shown on X at about 4pm JST on 27 August 2026: 9,337 likes, 1,206 reposts, 553 replies, 2,505,630 views.

The work is the system, not the prompt you typed this morning.

Official install paths are hermes update, or curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash. Read the Bot Mode docs before you point it at anything that can send, publish, or spend.

How does Grok Bot run the same idea?

Grok Bot is SpaceXAI’s hosted version of the same idea: named teammates you message, that work in parallel, and that come back when they need a decision.

The full definition is in what is Grok Bot. The short version, from the official 11 August 2026 launch and the product page: SpaceXAI hosts a persistent cloud computer. Your Bots sign into apps, keep working 24/7, and show up on official desktop and iOS. You message them like a teammate. You can teach a workflow by demonstration. No workflow builder required.

@bot widened access on 21 August (SuperGrok Plus, Cursor Pro+, Cursor Teams, plus a limited free trial), then again on 26 August (all SuperGrok and Cursor Pro, plus a weekly usage reset). The 26 August news post lists SuperGrok, SuperGrok Plus, SuperGrok Heavy, Cursor Pro, Cursor Pro+, Cursor Ultra, and Cursor Teams Standard and Premium. It has its own usage pool. Enterprise is a waitlist.

Entry prices on the official product page, checked 27 August 2026: Cursor from $20 a month, SuperGrok from $30 a month, Teams from $40 a seat a month.

The Grok Bot docs are blunt: all of a user’s Bots share one computer. Files, browser sessions, and logins are shared across the roster. Do not treat a second Bot as a lock.

Approvals are built in. Allow or Deny. Auto Review. Take control yourself for passwords, passkeys, two-factor codes, CAPTCHAs, and payment confirmations.

The real difference (isolation versus a shared computer)

On Hermes Agent, a Bot is a profile. Memory, skills, and chat history stay in that profile. Credentials do not: Bot Mode says new Bots share one OAuth/token pool with the main profile by default. Do not assume the researcher cannot see the finance Bot’s logins.

On Grok Bot, the computer is user-scoped. One persistent cloud machine. One set of files and browser sessions. Several Bots can have their own screen and work in parallel. They cannot pretend to be separate security boundaries.

If you care where a token lives, that sentence should decide the week. If you want a teammate that keeps going after you shut the lid, it should decide it the other way.

	Hermes Agent	Grok Bot
Runtime	You host it (desktop/gateway). MIT.	SpaceXAI hosts a cloud computer.
Isolation	Per-profile memory, skills, and history. Credentials share one OAuth/token pool by default.	One user-scoped computer. Logins and files are shared across the roster.
Models	Any provider/model Hermes knows, including local. Pin per Bot.	The SpaceXAI / Cursor / Grok stack that comes with the plan.
Surfaces	Desktop + messaging gateways (Slack, Telegram, etc.) + CLI.	Official desktop + iOS app.
Start cost	Software is free. You pay models, hardware, your time.	Bundled with an eligible paid SuperGrok or Cursor plan.
Approvals	You own the machine. Supervise like any agent with real system access.	Built-in Allow/Deny, Auto Review, take-control for passwords.
Fit	Technical founder, wants isolation, local models, or to keep data on their own machine.	Founder who wants a teammate that keeps working when the laptop is closed, inside the tools they already pay for.

I am not picking a winner. The choice is isolation plus control versus a hosted computer plus less ops.

Models, cost, and who does the ops

Hermes Agent talks to any provider and model it already knows. ChatGPT, Claude, a local model, a Nous endpoint. Pin a cheap model to the tidy-up Bot and a stronger one to research.

Grok Bot runs on the SpaceXAI / Cursor / Grok stack that comes with the plan. You are not standing up providers.

Hermes Agent the software is free. The invoice is models, a machine that stays up, and the hours you spend operating it.

Grok Bot is bundled with an eligible SuperGrok or Cursor plan. The extra software licence is not the decision. The decision is a hosted computer with a separate usage pool, and one shared login surface for the whole roster.

On Hermes Agent, you are the operator. If the machine sleeps, the Bot sleeps, unless you have put a gateway somewhere that does not. On Grok Bot, SpaceXAI is the operator. The laptop can close. The routine can still run. You still own the approval gate.

My honest read: pay for ops you will not do, or keep the data on a machine you will actually maintain. Do not pick Hermes Agent because open source sounds virtuous, then leave it on a laptop that goes into a bag at 5pm. Do not pick Grok Bot because it is already in the Cursor bill, then give every Bot the same production login and hope the names keep them apart.

What an Australian founder should actually do this month

Do not stand up a ten-bot company. Pick one marketing loop you already do every week. Follow-up drafts. A Monday research brief. A competitor check. A newsletter assembly. Recurring, messy, reviewable, valuable. AI can assemble the work. You still remain the judge of anything that gets sent, published, or spent.

If you already pay for Cursor or SuperGrok, and you want that loop running on a client lunch, use hosted Grok Bot. Create one named teammate. Message it like a person. Teach the path once if the demo is there. Put the send behind approval.

If you want the same loop on your machine, with isolation, local models, or data that stays on your machine, use Hermes Agent. Install it. Make one Bot. Write the SOUL.md. Pin the model. Give it only the skills that job needs. Watch hermes cron list.

That is the whole month. Systems over prompts. One named loop that still works on Thursday.

If you outgrow it, add the second Bot on purpose. On Hermes Agent, that second Bot is a separate profile for memory and skills. It is not a separate credential vault unless you stop sharing the token pool. On Grok Bot, it is another teammate on the same computer. Plan the logins accordingly.

FAQ

Is Hermes Agent free?

The software is free. Hermes Agent is MIT-licensed on GitHub. You pay for models (local or paid APIs), the hardware that stays on, and the time you spend operating it.

What can Hermes Agent do?

Hermes Agent can run a named roster of bots on a machine you host. Official surfaces are the desktop app on macOS, Windows and Linux, plus Telegram, Discord, Slack, WhatsApp, Signal, email and a CLI. Bot Mode (v0.20.3) gives each Bot its own role, model, memory, skills and credential configuration. New Bots share the main profile’s OAuth/token pool by default unless you deliberately reconfigure that boundary. Bots can message each other, sit in group chats, and run scheduled routines. You pick the models, including local ones. The software is MIT-licensed. You do the ops.

Can Hermes Agent use ChatGPT, Claude, or a local model?

Yes. Official Bot Mode lets you pin any provider and model pair Hermes already knows, including local. Different Hermes agents on the same roster can run different models side by side.

Is Grok Bot open source?

No. Grok Bot is a hosted SpaceXAI product. You get it with an eligible paid SuperGrok or Cursor plan. There is no MIT repo you can fork onto your NAS.

Which is better for a 10-person Australian team?

Neither is better as a mascot. A ten-person team needs shared work, a clear approval rule, and someone who owns the system when it breaks.

If the team already lives in Cursor, Cursor Teams includes Grok Bot (from $40 a seat a month on the official page, checked 27 August 2026). That is the less-ops path: one hosted computer per user, desktop plus iOS, built-in approvals. Enterprise-wide rollout is still a waitlist.

If the team wants isolation, local models, or data on their own machines, Hermes Agent can do the roster. It will not give you a hosted enterprise admin suite. Official docs do include an optional dashboard with OIDC and a login audit file. Someone technical still has to own the runtime.

Sources I checked (27 August 2026)

• Hermes Agent site: hermes-agent.nousresearch.com
• Bot Mode docs: docs/user-guide/bot-mode
• Hermes Agent GitHub (MIT licence, v0.20.3 / v2026.8.16.2): NousResearch/hermes-agent
• @NousResearch Bot Mode (17 August 2026)
• @bot launch (11 August 2026)
• @bot access, 21 August and 26 August
• Grok Bot launch, 11 August 2026: x.ai/news/introducing-grok-bot
• Grok Bot product: x.ai/bot
• Grok Bot overview: docs.x.ai/grok-bot/overview
• More plans, 26 August 2026: x.ai/news/grok-bot-more-plans

If you want a wider view of the system around the loop, the Digital Bond scorecard takes a few minutes and does not ask for an email. If you would rather run these systems with other founders instead of collecting tutorials, AI Systems Lab is the room for that.

I still care which tool you pick. I care more that one named loop is actually running on Thursday.