THE SHORT VERSION
Hermes only became practical when we stopped treating it like a clever chat window.
For us, the better question wasn’t “can AI answer a prompt?”
It was “can AI sit close enough to the work to prepare the next step, without being allowed to make a mess?”
That’s where Hermes got interesting. It gave us a way to connect AI to tools, memory, skills, scheduled jobs and team chat, while still keeping approval and judgement with people.
WHY WE CARED
The normal chatbot pattern wasn’t enough.
I’ve tested a lot of AI tools. Some genuinely save time. Some are toys. Some feel like another tab you have to open before copying the answer back into the real work.
That’s fine for a one-off draft. It’s not enough when the job needs context from HubSpot, Drive documents, Slack, meeting notes, project history, campaign tools and the way CLCK actually works.
What we wanted was closer to a practical assistant inside the business. Not a public chatbot. Not something replacing the team. More like a capable operator that can gather context, draft the next step and show its work, while a person still decides what happens next.
That’s why we started paying attention to Hermes. The Nous Research GitHub repo showed an agent built for more than a single chat window. OpenRouter’s public app and agent rankings were another signal that people were using it. I’d treat that as a signal, not proof that every business should copy it.
PLAIN ENGLISH
Hermes is the layer around the model.
A simple way to think about it: the model does the thinking and writing. Hermes gives that model a place to work from, a memory, reusable procedures, approved tools and a way to talk to the team.
So the business question isn’t “which chatbot is smartest?” It’s “what work are we comfortable letting AI prepare, what sources can it use, and where does a person approve the output?”
If you want the technical detail later, the Hermes docs are the place to go. I’d read this story first, because the harder question usually isn’t installation. It’s where AI is allowed to help.
Sales call summaries
Turn notes, transcripts and CRM context into a short follow-up brief a human can check.
Proposal prep
Gather scope, exclusions, similar work and open questions before someone writes the commercial version.
CRM hygiene
Flag stale deals, missing next steps and unclear owners so the team can clean up the important ones.
Client meeting prep
Pull the original goals back into view, not just the latest email thread.
Campaign reply triage
Classify replies, surface warm leads and keep any external response behind review.
Internal knowledge support
Find the right source, draft a clear answer and turn repeated answers into a reusable procedure.
WHAT WE TRIED
Our path wasn’t neat. It got clearer as the work got real.
This is the simple version of the CLCK journey: start with AI as a power tool, then gradually move it closer to real operating work.
- 1
Early AI and CLI work
Good for building, testing and digging through files, but too technical for most team requests.
- 2
Slack for quick help
Slack helped because the team already worked there. It got messier once the work needed scope, evidence and review.
- 3
Discord trial
Worth trying, but it never felt like the right home for CLCK operations.
- 4
Zulip topic lanes
Topics gave us cleaner lanes for strategy, execution, review and follow-up.
- 5
Google Cloud VM
A stable business home, instead of a laptop that had to be open at the right time.
- 6
Parent and child workflow
One lane decides the work. Another completes a bounded slice and comes back with evidence.
The pattern that emerged was pretty simple. The more important the work became, the more structure we needed around it.
Quick chat was fine for quick questions. Once Hermes was helping with work that needed source material, review and a final decision, the workspace mattered just as much as the model.
WHERE IT LIVES
Slack helped. Discord didn’t stick. Zulip gave us lanes.
The platform choice sounds like a small detail until you’re trying to review what the agent actually did.
Slack
Great front door for quick work.
Slack helped because CLCK already used it. Quick summaries, lookups and approvals felt natural there. The problem was long structured work. A thread can become strategy, execution, source material, debate and final answer all at once.
Discord
Worth testing, not our operations home.
Discord made sense to try because Hermes and agent communities often live there. For CLCK’s day-to-day work, it didn’t become the place we wanted client, sales and ops decisions to live.
Zulip
Better for structured work lanes.
Zulip’s topics gave us a cleaner way to separate a parent decision lane from child execution lanes. It takes discipline, but serious agent work needs the work to stay findable and reviewable.
Good for a fast ask or approval.
Worth testing, but not where CLCK wanted operations to settle.
Better when the work needs a clear lane and review trail.
HOSTING
A laptop was fine for learning. A VM made more sense for business work.
Early experiments can live on someone’s computer. That’s normal. But once Hermes became part of CLCK’s operating rhythm, relying on a laptop felt wrong.
We moved to a Google Cloud virtual machine because the agent needed a stable home for chat, tools, scheduled jobs and integrations. If you’re comparing options, Google’s Compute Engine VM documentation explains what a VM is. The business issue is simpler: who owns the thing when it becomes important?
Slack, Zulip, files, CRM, calendars and campaign systems
Gateway, scheduled jobs, skills, memory and approved tool access
OpenRouter or direct model providers, chosen for the job
Why it made sense for us
A VM gives the operator a stable private place to run. It’s there for scheduled work and team requests, not just when someone has a laptop open.
The trade-off
Someone has to own updates, logs, secrets, access, backups, monitoring and cost. A cloud server isn’t a magic safety blanket.
KEEPING WORK BOUNDED
Parent and child became our way to keep the agent from turning work into soup.
One of the easiest ways to lose control with AI agents is to keep everything in one giant thread. Strategy, source material, edits, tool output, questions, mistakes and final answers all blend together.
Our parent and child pattern fixed a lot of that. The parent lane decides the work and reviews the result. The child lane gets one bounded job, does the inspection or build, verifies it, then reports back in a compact format.
Damien or parent lane
Decides the slice, source of truth and boundaries.
Bounded child task
Inspects, builds or researches one clear piece of work.
Verification
Runs checks and reports what changed, what passed and what’s risky.
Decision
A person chooses the next step before anything sensitive goes live.
Safe to prepare
Summaries, checklists, context packs, draft options, missing-field reviews and next-step suggestions.
Needs a person first
External emails, public posts, CRM changes, client-facing copy, campaign settings, quotes and commercial recommendations.
Leave until later
Deleting data, changing live campaigns, sending messages unsupervised or giving broad access before the workflow has proved itself.
START HERE
Don’t ask AI to run a department. Give it one clear job.
If the job is already worth doing manually, it might be worth teaching an agent to help. If nobody understands the job yet, automate nothing.
- 1
One repeated job the team already understands
- 2
Known source material the agent can inspect
- 3
A clear output, like a brief, list, draft or decision pack
- 4
One approval point before anything external changes
- 5
A safe next step if the answer is uncertain
Good first jobs are usually boring: summarise a sales call, prepare proposal context, check stale deals, draft a follow-up, or turn a messy client thread into a short decision pack.
The first win isn’t the AI doing something dramatic. It’s taking a job that normally falls between people and making the next step obvious.
WHAT I’D DO DIFFERENTLY
If we started again, I’d make the operating rules clearer sooner.
DO EARLIER
Name the first job earlier
It’s tempting to ask the agent to help with everything. The better move is choosing one repeated job, then making that job boringly clear.
Decide where work should live earlier
The chat surface isn’t cosmetic. It changes whether work stays visible, searchable and reviewable.
Write approval rules before the first live use
The question isn’t whether the AI is smart. It’s what it’s allowed to do before a person checks it.
Treat hosting as an operations choice
Once an agent is part of the team rhythm, uptime, logs, backups, updates and recovery matter.
SKIP OR DELAY
I wouldn’t start with a vague department brief
“Help with sales” is too broad. “Prepare a follow-up pack after a discovery call” is much easier to test.
I wouldn’t over-explain the tech to the team
Most people don’t need to care about the runtime. They need to know what to ask, where the answer appears and when they’re responsible for approval.
I wouldn’t save every correction forever
Memory and skills are powerful when they stay clean. They become a liability if every temporary detail gets treated as permanent truth.
I wouldn’t connect everything on day one
Read-only, narrow and reviewable is a much safer starting point than a giant access project.
WHERE TO LOOK NEXT
Read the technical docs after you’ve picked the business job.
If you want to go deeper after this, start with the links below. Just don’t let the setup detail distract from the bigger decision: where is AI allowed to help, and who approves the work before it matters?
If you’re comparing model access routes, the OpenRouter Hermes integration guide is worth reading. For a business owner, the simple version is this: OpenRouter can make model access easier, while direct providers may suit different privacy, support or procurement needs.
CLCK SUPPORT
If this is starting to sound like operations, that’s what this becomes.
Hermes is the tool. The harder work is choosing the first business job, setting the boundaries, deciding where the work lives, and making sure the team can review what happened.
CLCK can help you work through that first operating lane before you connect AI to systems that matter.
TALK THROUGH A FIRST WORKFLOWFAQ
Hermes Agent for business FAQs.
What is Hermes Agent in plain English?
Hermes Agent is an open-source AI agent from Nous Research. In business terms, it gives an AI model a more practical operating layer: tools, memory, skills, scheduled work and messaging surfaces.
Is this an official Hermes guide?
No. This is CLCK’s lived-experience guide. We don’t own Hermes and this isn’t official support. This is what we’ve learnt using it inside CLCK.
Why not just use ChatGPT or Claude?
A normal chatbot can help with one prompt at a time. The gap we were trying to close was different: context, repeatable workflows, approvals and work that sits across files, CRM, team chat and scheduled jobs.
Why did CLCK move serious agent work into Zulip?
Slack still worked for quick tasks, but long threads were hard to review. Zulip topics gave us clearer lanes for parent decisions, child execution, summaries and follow-up.
Why run Hermes on a Google Cloud VM?
A laptop is fine for learning. For CLCK, Hermes needed a stable place to run gateway, scheduled and integration work, so a Google Cloud VM made more sense. That also means someone has to own updates, access, logs, costs and recovery.
What should a business try first?
Pick one repeated job you already understand, such as summarising a sales call, preparing proposal context, checking stale deals, drafting a follow-up or creating a client meeting brief. Keep the first version narrow and reviewed.
Can Hermes use OpenRouter?
Yes. OpenRouter has Hermes integration documentation, and it can be a simple way to access multiple models. The business decision is whether that route suits your privacy, reliability, cost and support needs.
Is Hermes Agent open-source?
The public Hermes Agent repository shows an MIT licence. If that matters for your use case, check the current licence file and get proper advice for anything legal or high-risk.