SME Centre@SMF  ·  Business Advisor session

Hermes Agent
and Meetily

An assistant that runs on your own machine, and a meeting note-taker that never uploads your audio. Two tools, and they take back the half of the job nobody became an advisor to do.

Lewis Ang  ·  1 October 2026  ·  30 minutes, no live demo
01
The problem

Our knowledge sits in silos, and the shared record is written too late.

How a case actually runs

  • The only shared record of a case is what one advisor typed into PEP, and it may not capture the SME's real interests, issues or the solution needed.
  • A client can meet a different advisor from the same Centre next session. Until someone opens PEP, the new advisor starts cold.
  • PEP is usually read after the meeting, not before, so the preparation happens backwards.
  • The write-up is tedious and repetitive, and it is written from scratch every time.

What it costs the Centre

  • Knowledge lives with individuals, so it degrades on leave, handover and turnover.
  • Nobody can quickly ask what hundreds of clients asked for, so workshop themes and group upgrading projects run on hunch.
  • With no consolidated view of interests and pain points, workshop invitations are guesswork and events pay for it in no-shows.
  • There is no AI SOP, so advisors already use personal chatbots on confidential material.
02
The idea

A chat window helps one advisor. A shared assistant helps the Centre.

A personal chatbot answers you, in one tab, and keeps nothing the Centre can find again. What it learns stays with you, and the case record stays whatever PEP says.

An assistant is the other shape. It sits on a machine, keeps a record the whole team can read, and keeps working while you are in a meeting.

  • One named bot per advisor, all reading and writing the same Centre record, so what one of us learns the rest can ask for.
  • The meeting summary is drafted into the PEP format for you to check and paste, instead of being written from scratch.
  • Meetings are transcribed and filed on the machine, so the audio and the minutes never leave it.
  • Mid-meeting: what are other companies in this sector doing, and what have they applied for, answered from our own records.
  • It acts only on rules the Centre sets. Nothing reaches a client without one.
03
The two tools

One listens. One thinks. Neither forgets.

The assistant

Hermes Agent

  • Free and open source. Runs on the Centre's own machine.
  • Each advisor gets their own named bot, and every bot reads and writes the same shared record.
  • It answers from our case corpus: companies, contacts, interests, pain points, what has been applied for.
  • You talk to it from Telegram, Slack or WhatsApp.
  • Plain notes on disk, so any of us can read, correct or delete it without a vendor.
The ears

Meetily

  • Free and open source. Transcribes meetings on the machine.
  • No bot joins the call, so the client sees nothing unusual.
  • Start it before you join. The audio never leaves the laptop, so there is nothing to leak.
  • It drafts the summary that becomes the PEP write-up and the minutes, instead of writing either from scratch.
  • Works with Zoom, Teams, Meet and Webex, because it listens to system audio.
Together: one record the Centre owns, and the writing around every meeting done for you.
04
The assistants you already know

Four good products. One catch.

OpenAI

ChatGPT

The one everybody has open.

  • Friendliest of all
  • OpenAI models only
  • Cannot see your folders
  • Free, then US$20/month
OpenAI

ChatGPT Work

ChatGPT sold to organisations.

  • Admin controls, central billing
  • OpenAI models only
  • US$25 monthly or US$20 annual per seat
  • Two seats minimum
Anthropic

Claude Cowork

Delegate a goal, come back to finished work.

  • Reads and writes your files
  • Claude models only
  • Pro US$20/month
  • Files live with your account
Google

Gemini Spark

A 24/7 agent inside Google's apps.

  • Works with Gmail, Drive, Docs
  • Gemini models only
  • Needs Google AI Pro, S$28.99/month
  • Personal accounts only, for now
All four rent you one company's models, per person, every month, for as long as you use them.
05
The three that break the pattern

Open source, your machine, any model.

Ready-made

Vellum

A personal assistant for Mac, with one memory across phone, web and chat apps.

  • Free base tier
  • Paid from US$30/month
  • Bring your own model keys
  • Open source, can be self-hosted
Do it yourself

OpenClaw

Self-hosted gateway that drops an assistant into the chat apps you already use.

  • Software free
  • You pay hosting and model use
  • Roughly US$11 to US$35/month typical
  • Any model provider
What we are showing

Hermes Agent

Runs on a Mac you already own, with the memory and the notes on that machine.

  • Software free
  • Model use only
  • Real bill here: US$10/month
  • 300+ models, swap any time
The difference is not intelligence. It is who owns the assistant, the data and the choice of model.
06
Assistants, side by side

The columns that actually decide it.

AssistantFriendlyCustomisableRuns onModel-agnosticCost, per user
ChatGPTHighestLowTheir cloudNoFree to US$20/mo
ChatGPT WorkHighLowTheir cloudNoUS$25/mo, or US$20 annual
Claude CoworkHighMedium to highDesktop + their cloudNoFrom US$20/mo
Gemini SparkHighMediumTheir cloudNoNeeds AI Pro, S$28.99/mo
VellumHighMedium to highYour Mac or theirsYesFree, then from US$30/mo
OpenClawLowestHighestYour machine or a serverYesFree + usage, about US$11-35/mo
Hermes AgentMediumHighestYour machineYesFree + usage, US$10/mo here
Full detail, with sources and dates, on the Assistants page of this site.
07
The new shape: bots

Not one assistant. A roster of named ones.

Bot familyRuns onModel-agnosticCost
OpenAI DotsTheir cloud, own computerNoNeeds ChatGPT Pro, US$100+/mo
Grok BotTheir cloud VMNoUS$20 to US$300/mo, no free tier
Hermes Bot ModeYour machineYes, per botFree + usage, US$10/mo here
OpenMausBotYour machineYes, via your CLIsFree, open source
  • A bot is not a prompt. It has a name, a job, its own memory and its own schedule.
  • Bots message each other and hand work between themselves.
  • You talk to them like colleagues, from a chat app.
  • The one question that matters: are their hands on your machine, or in someone else's cloud?
Full detail, with sources, on the Bots page of this site.
08
Meeting notes, compared

Where does the client's voice actually go?

ToolWhere the audio goesPriceWho said what
Transcribe (GovTech)Government cloud, cleared for sensitive materialFree to public officersSummaries and Q&A
MeetilyYour own laptop. Never uploaded.Free, or US$10/user/mo ProYes, on Pro
OpenWhisprYour own laptopFree, or US$8/moYes, free
Otter.aiTheir cloud, US companyUS$8.33 to US$30/seatYes
Fireflies.aiTheir cloud, US companyUS$10 to US$39/seatYes
HappyScribeTheir cloudFrom US$29/mo, human from US$2/minAutomatic
Full detail, including the eligibility catch on the government tool, on the Transcription page of this site.
09
What it costs

Ten US dollars a month. That is the whole bill.

AI model usageUS$10 a month on the current plan, which itself includes US$70 of model credit.
Web searchCovered by a free monthly credit, then US$5 per 1,000 searches.
Everything elseNothing. Telegram, Slack, the assistant software and the free tier of the meeting transcriber all cost zero.
HardwareA Mac already owned, left switched on. Electricity only.
US$10 / month

About the price of one lunch a month, and there is no per-seat licence: the bill follows usage, not headcount.

Figures read from the provider's own pages on 1 October 2026, including the plan's published credit allowance.
10
Why the pricing model matters

Per seat forever, or per use.

The rented model

  • Every colleague needs their own seat, at US$20 to US$30 a month.
  • Ten advisors is US$2,400 to US$3,600 a year, and it grows with every new hire.
  • The money stops the day you stop paying, and nothing stays behind.
  • You use the vendor's model, at the vendor's price, for as long as they allow.

The owned model

  • One model bill, whether one advisor uses it or ten.
  • The marginal cost of the tenth advisor is close to zero.
  • The memory and the client record stay on the machine, in a folder we can read.
  • When a better or cheaper model appears, you switch, and nobody has to relearn anything.
This is the barrier worth breaking: not the AI, the rental.
11
Privacy, in one slide

Nearly nothing leaves the machine. That is the point.

  • Client meetings are transcribed on your own laptop; the audio is never uploaded.
  • The assistant's memory is a folder of notes on the machine, which any human can open and read.
  • The only things that go out are the question you type, the search request, and the finished summary if you chose a cloud model to write it.
  • Client documents are a processing input, not an archive: extract what is needed, then delete the source.
Why it matters here

PDPA, briefly

Personal data you send to a vendor's cloud is still your organisation's responsibility: consent, notice of purpose, retention, overseas transfer and a breach clock that starts the moment you know.

Keeping client material on the machine removes most of those questions instead of managing them.

12
Setup, part 1

Hermes Agent, in three commands.

  1. Install it. Download the desktop app, or paste this into Terminal:
    curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
  2. Reload your shell:
    source ~/.zshrc
  3. Choose your model, then check your health, then say hello:
    hermes model   hermes doctor   hermes
Needs an Apple Silicon Mac. The full walk-through, including the one-command portal setup, is on the Setup guides page.
13
Setup, part 2

Telegram and Slack, without opening a port.

Telegram, 60 seconds

Scan and go

  • In the assistant's app, open Messaging, then Telegram.
  • Click Create with QR and scan it in Telegram.
  • It creates the bot, saves your user ID, and restarts itself.

Manual route: message @BotFather, send /newbot, name it, and copy the token it gives you.

Slack, 5 minutes

Socket Mode, no public address

  • Run hermes slack manifest --agent-view --write
  • Paste that file into a new app at api.slack.com/apps.
  • Enable Socket Mode, allow direct messages, install to your workspace.
  • Paste the three tokens into the assistant's settings file and restart.
Socket Mode means the bot dials out to Slack. Nothing on your laptop is exposed to the internet.
14
Setup, part 3

Meetily, in one sitting.

  1. Download it from meetily.ai, open the file, and drag Meetily into Applications.
  2. On first launch, allow the microphone and system audio. Without system audio it cannot hear the other side of the call.
  3. Download a local transcription model when asked, and choose local summaries (Ollama) if you want nothing at all to leave the machine.
  4. Start the recording before you join the call. No bot joins, so the client sees nothing unusual.
  5. After the call, generate the summary and export it. Markdown is free; PDF and DOCX come with Pro.
Speaker labels, custom templates and the connection to your assistant need Pro: US$10 per user per month, billed annually.
15
Day one

Five things to try before the next session.

  • Ask it to summarise a document you were going to read twice.
  • Have it draft the meeting confirmation you send anyway.
  • Record one call in Meetily and read the summary afterwards.
  • Ask it a question about your own files, and see what it can already reach.
  • Turn one repeating task into a schedule, and see if it holds for a week.
None of these five needs a client name or a client document.
16
Honest limits

What it will not do.

  • It is not approved software yet. This session is how we find out what the Centre is willing to allow, and on whose machine.
  • It needs a machine left switched on, and a home internet connection that stays up.
  • You have to be willing to open a terminal once, and to read the error if something fails.
  • Meeting notes are only as good as the audio and the model. A bad recording produces confident nonsense.
  • It is not a decision-maker. It drafts, it reports, it reminds. You decide, and you send.
17
Where this could go here

The part that is bigger than one advisor.

Client knowledge, held once

  • One searchable record of what we know about each company, rather than eight private versions of it.
  • Before a call: "what do we know about this company", answered in seconds.
  • Every advisor's questions sharpen the same record, so the tenth advisor starts where the ninth finished.

The Centre's own work

  • Map what clients actually ask for, by industry and by pain point.
  • Curate events and workshops from evidence rather than hunch.
  • Group-based upgrading gets one shared evidence base instead of scattered spreadsheets.
18
Take it away

Three things, and one rule.

  • The barrier is not the AI. It is the rental: you can own the assistant and choose the model.
  • The safest meeting notes are the ones that never leave your laptop.
  • Start with one task, not a platform. Ten minutes a day is enough to know if it earns its place.
One rule that does not bend

Client information stays with the client

It is never posted into a group channel, and the assistant is instructed never to put it anywhere other people can see it. That rule is the reason the rest of this works.

Questions welcome.  ·  Lewis Ang  ·  1 October 2026
19
Arrow keys or the buttons move through the deck  ·  press P to print or save as PDF
Overview Assistants Bots Transcription Setup guides 1 / 18