An agent on a machine in the office, a model that never leaves it, and ClickUp finally staying up to date. What it does alone, what we forbid it to do, and the rules that make the whole thing workable.
Hermes is an AI agent that runs on a Mac in the office with a local model and works in ClickUp like a member of the team: it turns support emails into tickets, summarises, chases, updates records and acts when it is mentioned. Sensitive replies and quotes stay human.
Why a local agent rather than ClickUp’s built-in AI
ClickUp is our reference tool: client tickets, projects, time spent. Its problem is never a missing feature, it is the discipline it demands. A forgotten follow-up, a status left on “in progress” for three weeks. The agent described here does nothing spectacular: it does the thankless part, every time.
Few SMEs are there yet. According to the data published by Eurostat on the use of AI in enterprises, extracted in December 2025, 19.95% of European enterprises with ten or more employees used AI in 2025, up from 13.48% a year earlier, and only 17% of small enterprises against 55.03% of large ones. Among those that use it, 31.05% use it to organise administrative or management processes, which is ClickUp’s territory.
ClickUp offers its own AI: according to its official documentation (in French), it is an option on the paid plans, billed per member per month, that gives access to models such as ChatGPT or Claude without leaving the tool. It is often the right choice.
We took another route for two reasons. Our tickets contain website logins, incidents in progress and billing details, and we did not want them to pass through an AI provider as well. Processing that stays in the office adds no supplier to manage, which simplifies the documentation the CNPD expects. The second reason is economic: a per-member option adds up every month, whereas a machine we already own costs only electricity. We have no measured figure for the saving.
Hermes in practice: what runs in the office
Hermes is not software we wrote. It is Hermes Agent, an open source project under the MIT licence published by Nous Research. Its official documentation, consulted in September 2026, describes it as an agent that learns as it works: it creates skills from what it does, keeps a memory from one session to the next, ships with a scheduler, connects to more than twenty messaging apps and speaks the MCP protocol to plug in third-party tools. We installed it as it is, then configured it.
The model doing the thinking behind it is Qwen 3.6, 35 billion parameters with a mixture of experts, served by Ollama on a Mac in the office. This open-weight model, under the Apache 2.0 licence, activates only a fraction of its parameters for each word it produces, which makes it workable on a single machine. No request leaves that machine.
The connection comes down to three pieces: a ClickUp account of its own for Hermes, invited with edit rights on our ticket lists; an MCP server to read and write; and a webhook that alerts it as soon as a comment mentions it or a task is assigned to it.
The five tasks Hermes does for us
| Task | Trigger | What Hermes produces | Who stays in charge |
|---|---|---|---|
| 1. Support emails into tickets | Every incoming email | Ticket in the right list, urgency, owner, acknowledgement | The ticket owner |
| 2. Summaries and reports | Fixed time | Overview of open tickets, delays, news | Everyone reads |
| 3. Follow-ups and escalations | Ticket with no movement | Reminder, escalation to a human, proposed closure | The owner, then management |
| 4. Statuses, fields and time | Action completed | Status, fields, tracked time, comment | Review at invoicing |
| 5. Action on request | Mention or assignment | Search, summary, progress, proposed reply | The person who mentioned it |
1. Support emails become tickets
Every email received in the support mailbox is sorted according to a table we wrote: website, print, administrative or development. Hermes sets the urgency, assigns the owner and sends an acknowledgement in the client’s language, naming the person in charge and giving a turnaround we can keep.
2. The morning update arrives without anyone writing it
The built-in scheduler sends reports at fixed times to the team’s messaging app: what has come in, what is dragging, what has been delivered. Their value lies in the fact that they exist every single day.
3. Follow-ups no longer depend on human memory
A ticket that nobody has picked up beyond the set deadline goes up to management. A ticket resolved two working days ago receives a closing message. If the client replies on a closed ticket, Hermes reopens the same ticket and alerts the owner.
4. Statuses, fields and time follow reality
With each action, Hermes updates the status, the fields and the tracked time, and leaves a comment saying what it did and why. End-of-month invoicing finally relies on accurate records.
5. Mention it and it acts
Mentioning Hermes in a comment or assigning it a task is enough for it to act: look up a piece of information, prepare a summary, suggest a reply, move a pending matter forward. It answers in the task thread, and the person who mentioned it keeps the decision.
The three tasks we keep human
An agent that does everything is an agent you end up unplugging. So we wrote down what Hermes does not do.
Replying to a client outside a validated template. Ambiguous email, unhappy client, threat to cancel, legal matter: the ticket is created, a person is alerted immediately, and no automatic reply goes out.
Making a technical diagnosis or taking on a commitment. Hermes never puts forward a probable cause for a fault, never promises a turnaround shorter than the standard one, and decides nothing on pricing, tax or law.
Handling a sales enquiry. A new need or a request for a quote is not a ticket: Hermes passes it on with its context to the person in charge of sales, and stops there. It is, in any case, never allowed to delete a task.
The rule of doubt. When a situation fits no rule, Hermes does not decide: it asks the question in a comment, mentions the owner and moves on. One question too many costs nothing.
The rules that make the agent workable
In a press release from June 2025, Gartner predicts that more than 40% of agentic AI projects will be abandoned by the end of 2027, because of runaway costs, unclear business value or inadequate risk controls. Our answer is not technical, it is procedural.
| Rule | What it requires | What it avoids |
|---|---|---|
| Kill switch | A “pause” label on a control task stops it on its next run | Cutting the machine off in an emergency |
| Autonomy levels | Supervised: every outgoing message waits as a draft. Autonomous: sent directly, from validated templates only | An improvised text sent to a client |
| Message identifier | Every email processed leaves its identifier on the ticket | Duplicates when a run is replayed |
| Write scope | Ticket lists, the prospect list and the log, nothing else | A change elsewhere in the workspace |
| Signed log | One dated line per run: actions, targets, result | No longer knowing who did what |
Hermes starts in supervised mode: its messages wait as drafts in the ticket until explicitly approved. Autonomous mode is earned after a period with no corrections, and lost the same way. Every human correction is recorded in its log.
Where to start in your SME
The right question is not “which agent should I choose”, it is “which task should I give it first”. Your requests must arrive in one known place: an agent cannot sort what it cannot see. Then choose a repetitive, low-risk task, such as the acknowledgement and the sorting. Finally, name the person who validates the outputs during the supervised period: without a validator, there is no trust.
For a Luxembourg SME, funding changes the decision. The SME Packages AI scheme described on guichet.lu (in French), page updated in March 2025, covers the integration of AI tools into the company’s processes at 70% of eligible costs, for a project between 3,000 and 25,000 EUR excl. VAT, which means a maximum grant of 17,500 EUR, reserved for SMEs established in Luxembourg.
An agent running in the office changes nothing for a cross-border commuter colleague working from Belgium or France: it is the machine, not the employee, that processes the data. We go into this logic in our article on local AI in business, and into choosing the tool in the one on the CRM that stays empty. Our team automates this kind of process for SMEs in Luxembourg and the Greater Region.
Frequently asked questions
Do you need an expensive machine to run a local agent?
No. The model we use runs on a recent laptop Mac with 32 GB of unified memory. Choose the tasks and the model first, then size the machine, not the other way round.
Does Hermes replace ClickUp’s built-in AI?
No, it answers a different trade-off. The built-in AI runs in the cloud and is paid per member. A local agent runs on your own hardware and sends no data out to think. The two can coexist.
What happens when the agent gets it wrong?
In supervised mode, the error stays in a draft that nobody has sent. In autonomous mode, it can only send texts from validated templates. In both cases, the kill switch stops it on its next run.
Do your clients know an agent is writing to them?
Yes. Every acknowledgement states that it was prepared by an automated assistant and names the person who will handle the request.
Is a local agent GDPR compliant by design?
It simplifies compliance without guaranteeing it: no extra supplier to manage, but you remain responsible for the record of processing activities, for informing individuals and for retention periods. We are not lawyers, so have a sensitive processing activity validated by your legal adviser.
Can you do the same with Notion, Jira or Trello?
Yes, as soon as the tool exposes an official interface and, ideally, an MCP server. The value lies in the written rules, the routing table, the validated templates and the autonomy levels.
Sources
- Eurostat, Use of artificial intelligence in enterprises, data extracted in December 2025.
- ClickUp, Qu’est ce que l’IA de ClickUp (in French), official help centre, consulted on 11 September 2026.
- Nous Research, official Hermes Agent documentation, consulted on 11 September 2026.
- Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, press release of 25 June 2025.
- Guichet.lu, SME Packages AI (in French), page updated on 11 March 2025.









