Grok Bot for Business: How We Run It, and How It Compares with OpenClaw and Hermes

Give each bot one job, connect it only to the systems that job needs, write your company procedures once as shared skills, and make every bot stop at an approval before it sends, posts or pays. That is how we use Grok Bot at Osher Digital: one Chief of Staff bot that we talk to, seven specialist bots that report through it, and nothing external happens without a person saying yes. If you would rather host the agent yourself, OpenClaw and Hermes Agent are good open-source options, but someone on your team then owns the server.

We are Osher Digital, a Brisbane automation consultancy. We build AI agents, n8n automations and system integrations for Australian businesses, and we run our own back office on Grok Bot. This post is the setup as it stands in October 2026, what we would tell a client to copy, and where a self-hosted framework is the better choice. We have no commercial relationship with SpaceXAI or Cursor, the companies behind Grok Bot, and nobody paid for this post.

---

Why most first attempts stall

Two set-ups fail in the same few weeks.

One chat thread doing everything. The same assistant triages the inbox, checks invoices, audits ads and drafts proposals. Context bleeds: the firm tone meant for a vendor dispute turns up in a client proposal, and last month's figures end up in this month's quote. That one bot also holds write access to every system at once, so one misread instruction can reach anything.

A self-hosted agent that becomes a side job. OpenClaw and Hermes Agent are capable open-source projects, and both use portable SKILL.md skills and keep memory between sessions. In a business, though, someone has to own the server. When a Microsoft 365 token expires, a runtime update breaks the background process, or the browser session logs out, that someone is usually the founder. An agent that reads untrusted email and has shell access and API keys on the same machine is also an easy target for prompt injection unless its permissions are locked down.

A managed workspace removes the server work. Separate bots with one job each remove most of the context bleed, and approvals limit what a mistake can do.

---

Our setup: one front door, seven specialists

We run eight work bots. The Chief of Staff, which we named Lucinda, is the only one we talk to day to day. The seven specialists stay quiet unless something needs a decision, and they send flags and drafts through Lucinda, not into their own chat threads. The one exception is a real outage in production or payments.

Bot	Job	Connected to	Routines and limits
Lucinda (Chief of Staff)	Front door and coordinator	Calendars, Granola meeting notes, its own mailbox, the other bots	Weekday morning brief, 7am calendar digest, a prep pack before each external meeting, a list of promises not yet kept, a weekly self-check
Chief	Founder's inbox and operations	Microsoft 365 mail and calendar, the hosting platform	Checks the inbox hourly in business hours. Never marks mail as read unless asked. Archives routine server notices. Supplies pricing and project history before calls
Sales	CRM and outbound pipeline	Our CRM and the outbound prospecting tool	Checks every prospect list against a do-not-contact list before spending credits. A daily check on active sequences
Developer	Scoping and delivery	SharePoint, the CRM projects module, GitHub, Granola	Scopes builds with Sales, creates each project's folders, tracks milestones, writes the handoff pack for engineers
Marketer	Search and paid ads	Search Console, GA4, Google Ads, Bing Webmaster Tools, the website CMS	Weekly search and AI-citation check. A read-only weekly Google Ads review. Any new campaign is left paused
Accounts	Bills, invoices and payments	The accounts mailbox, Xero, Stripe	Tracks supplier bills, drafts Xero invoices against project milestones, watches Stripe payouts and disputes. Never sends, pays, refunds or moves money without approval
Sentinel	Client support and hosting alerts	The support mailbox, server health alerts	Archives routine disk and server notices, passes real support tickets to Lucinda, drafts replies for approval
Sentry	Bug triage (hidden from the sidebar)	Sentry and GitHub	Wakes on new or returning production errors. Ignores repeats and flapping alerts. Keeps quiet overnight and on weekends unless it is a P0. Opens a GitHub issue or a fix pull request, never a push to main

A ninth bot handles the founder's personal email and calendar and has no Osher accounts connected, so personal admin stays out of work systems.

---

Functional bots plus a bot per client project

The eight bots above are horizontal: each owns a company-wide queue such as billing, the support mailbox or search rankings. Delivery, on the other hand, runs client by client. When one Developer bot carries eight live builds, the details of one client (their legacy system, the people involved, what the signed scope leaves out, where their files live) compete with the next.

So we are adding a project bot for each active engagement, named by the engagement code we already use everywhere: client, number, topic, such as ACME-01-PORTAL.

1. Only for real work. A project bot starts when a deal enters paid discovery (such as a Business Process Audit) or is won. Hosting-only clients and early leads stay with the functional bots.
2. Seeded with five things: the engagement code, the CRM deal and project, the SharePoint folder (01-project, 02-delivery, 03-comms, 04-finance), the code repository, and the signed price, deliverables and exclusions.
3. Day to day: after a client call it pulls the Granola notes, updates the CRM tasks, files the notes in 03-comms, refreshes the engineering handoff pack and drafts the follow-up email for review.
4. No second inbox watch. Project bots never check the founder's mailbox on their own timers. Chief and Lucinda read it once and pass client emails to the right project bot.
5. Hidden at handover. When the final milestone invoice is paid and the project is marked complete, the bot is hidden from the sidebar. Its history and files stay searchable for when the client comes back for engagement 02.

---

Write each procedure once, as a shared skill

The most useful thing we did was separate who a bot is (its role and its connected accounts) from how we do things (shared skills). Each procedure is written once as a skill, a Markdown file of instructions, and any bot loads it when the task calls for it. When the procedure changes, every bot picks up the change.

• Engagement codes. One code per job, used the same way in the CRM deal, the project, the SharePoint folder and the proposal footer, so sales, delivery and accounts always mean the same job.
• Client folders. One client-wide 00-context folder for contracts and contacts, and one folder per engagement with the same four subfolders.
• Audit proposals. A branded Word proposal for a Business Process Audit, built from our template using only figures from the CRM, emails and meeting notes.
• Fixed-price proposals. An HTML proposal printed to an A4 PDF in our house style.
• Engineering handoff packs. One folder with the brief, the source emails with their headers, extracted attachments, call notes and earlier specs, ready for a senior developer or a coding agent.
• Email voice. Before drafting, the bot reads two to four recent sent emails from the same mailbox to match the register, then removes stock AI phrasing. The draft still waits for a person to send it.
• Google Ads checklist. Confirm the account ID, switch off Search Partners on Search campaigns, and leave every new campaign paused for review.

---

Grok Bot vs OpenClaw vs Hermes Agent

Checked on 27 October 2026 against each project's own documentation and release notes (OpenClaw 2026.10.7, Hermes Agent 0.22.5 and the Grok Bot docs).

	OpenClaw	Hermes Agent	Grok Bot
What it is	Open-source (MIT) personal agent with a self-hosted Gateway, formerly Clawdbot and then Moltbot	Open-source (MIT) agent from Nous Research, with a CLI, a TUI and a messaging gateway	A managed agent workspace from SpaceXAI (formerly xAI), run through Cursor, in beta since August 2026
Who runs the server	You. Node 24 or later, installed as a background daemon	You. Python 3.11 to 3.14, run locally or on Docker, SSH, Modal, Daytona or Vercel Sandbox	The vendor. One managed cloud computer per user, shared by all your bots
Where you talk to it	33 chat channels, from WebChat to WhatsApp	The terminal, or chat apps through its gateway	Desktop apps for macOS, Windows and Linux, and mobile apps
Connecting business apps	Skills and API keys you configure yourself	The Tool Gateway and a catalogue of about 600 reviewed plugins, with credentials you manage	Plugins including Outlook, Teams, SharePoint, OneDrive, Gmail, Xero, Granola and Apollo, plus MCP servers and webhooks
More than one agent	Several isolated agents in one Gateway, routed by bindings	Profiles and subagents	Several bots, each with its own role and routines; a bot can be hidden from the sidebar
Skills	Follows the Agent Skills format	Compatible with the Agent Skills format, and writes and improves its own skills	One private skills library shared by all your bots
Safety controls	Tool policies you configure	Its security policy says "the only security boundary against an adversarial LLM is the operating system"	Approval cards, Auto-review, outside content marked as untrusted data, and a secrets vault
Cost	Free software. You pay for the server and the model	Free software. You pay for the server and the model	Included with Cursor Premium (US$30 a month) and higher plans

OpenClaw's security record is the main thing to weigh. CVE-2026-25253, a one-click remote code execution bug, was fixed on 30 January 2026. In February, Koi Security found 341 malicious skills among 2,857 on ClawHub, and SecurityScorecard counted 40,214 OpenClaw instances exposed to the internet on 9 February. None of that makes OpenClaw unusable. It shows what self-hosting asks of you: keep it patched, keep the Gateway off the open internet, and vet every skill before you install it.

Grok Bot has a limit worth knowing too. All your bots share one cloud computer: the same files, the same sign-ins and the same browser sessions. Each bot has its own connected accounts, so a bot only reaches the plugins you give it. Anything signed in on the shared computer and its browser, though, every bot can reach, so separate bots are not a full security boundary. Keep browser sign-ins to tools every bot may use, and let approvals do the rest.

When self-hosting is the better choice: you have a developer who wants full control, the data has to stay on your own infrastructure (on Australian servers, for example), you need a chat channel Grok Bot does not offer, or you want to pick the model provider yourself. For a business that wants the work done without running the server, a managed workspace is less to maintain.

---

Three rules we would not run without

Quiet unless something is off. If an hourly inbox check messages you to say it read 14 emails and all is well, you will stop reading it within a week. Every routine we run ends with no message at all when nothing needs a decision. When Lucinda does speak, a lead has arrived, a deploy has failed, a bounce rate has passed 2% or a document needs sign-off.

Draft and stage, never send. No bot sends an external email, posts a message, launches a campaign or moves money on its own:

• Sentinel and Chief draft replies and wait for approval.
• Marketer builds new Google Ads campaigns in Paused status.
• Sentry opens a pull request; it never pushes to main.
• Accounts creates Xero invoices as drafts.

Hard limits on suppression and money. Before the Sales bot spends an enrichment credit or stages a sequence, it checks every prospect against one do-not-contact list built from active CRM clients, every domain we have ever invoiced in Xero and every open deal. The outbound tool will not do that check against your accounting system for you.

We build client systems the same way. The Affinity to Notion sync we built for Sharrock Pitman Legal keeps about 1,100 open matters in step with a practice management system holding more than 22,000. Every writer is off until it is deliberately switched on, guards refuse a run that looks wrong, and 341 automated tests cover the service. Reading can be broad. Writing has to be gated.

---

How to roll it out in four steps

1. Start with a Chief of Staff bot. Connect your calendar, your work email and your meeting recorder. Set up two routines: a 7am calendar digest and a prep brief before every external meeting that pulls recent emails and past call notes.
2. Write your top three procedures as skills. Before adding bots, write down how you name projects, how you format proposals and how you write emails. Every bot you add later inherits them.
3. Split out the noisy queues. Once the Chief of Staff is working, add bots for the queues that clutter your day: accounts, support, CRM and marketing. Make them report through the Chief of Staff.
4. Add project bots for active client work. Start one when a major project kicks off, point it at that client's folder, CRM deal and code, and hide it when the final invoice is paid.

---

Want this built around your own systems?

Setting up a Chief of Staff bot for one person takes an afternoon. Connecting agents to the systems your business runs on (Xero, Microsoft 365, SharePoint, your practice management or job management software, or a database nobody has documented) takes engineering, and the guardrails matter more than the prompts.

• Healthcare: we automated patient referral intake for Statera Health, cutting data entry time by more than 90%.
• NDIS: an AI agent now drafts property damage insurance claims for Strive Community Care, with about 70% less manual data entry.
• Legal: a guarded Affinity to Notion sync for a Melbourne law firm that has never overwritten a lawyer's notes.
• Property inspections: a Business Process Audit for a 35-person inspection firm found more than 8,000 admin hours and $250,000 a year to recover.

We start with a paid, fixed-price Business Process Audit: a day onsite following the work from first enquiry to money in the bank, then a build-ready report you own. Or book a 15-minute call to talk it through.

---

Frequently asked questions

Why use Grok Bot instead of OpenClaw or Hermes Agent?

OpenClaw and Hermes Agent are open-source agents you host yourself, on a VPS, a spare Mac or in Docker. That suits a technical owner who wants full control. For most businesses it means someone has to patch the server, register OAuth apps, rotate API keys and fix broken browser sessions. Grok Bot is a managed workspace with built-in connectors, a managed cloud computer and browser, and approvals before outbound actions.

How many bots should a business run?

Start with one Chief of Staff bot and add specialists only when a queue is noisy enough to deserve its own bot. We run eight: the Chief of Staff plus inbox and operations, sales and CRM, accounts, marketing, support, delivery and bug triage. Agencies and consultancies can add a bot per active client project and hide it when the project ends.

Can Grok Bot work with software that has no API or connector?

Partly. Your bots share a managed cloud computer with a browser. Once you sign in to a web-only tool there, the session usually stays signed in for later runs, although some sites expire sessions. We use that for Search Console, Google Ads and SharePoint. It works best for reading reports and routine admin. For anything that writes business records at volume, a proper integration is safer.

How do you stop a bot sending an embarrassing email or making changes?

Two ways. Grok Bot asks for approval before externally visible or irreversible actions, so a bot can read the inbox and draft a reply but cannot send it until you approve. And each routine is set up read-only by default: new ad campaigns are created paused and Xero invoices are created as drafts.

Can the same SKILL.md skills be used in Grok Bot, Claude Code, OpenClaw and Hermes?

Mostly. Claude Code, OpenClaw and Hermes Agent all follow the open Agent Skills format, so a skill's name, description and instructions carry across. Product-specific extras do not. Grok Bot keeps its own skills library and does not document importing SKILL.md files, so plan to copy the instructions across by hand. Because a skill is plain Markdown, that takes minutes.