Most “AI assistant” pitches are a chat box with a monthly fee. This is the other kind: a piece of software that runs on its own server, on a schedule, with tools — and gets cheaper the more boring the work you give it. The parts cost about the same as one premium AI subscription, and we know they work because this publication runs on exactly this stack.
The combination is the open-source Hermes agent and the open-weight model MiniMax-M3, paid for through Nous Portal’s $20-a-month Plus plan, sitting on a small virtual server.
The stack, in pounds
The numbers, grounded: Nous Portal Plus is $20/month and bundles 300+ models — including MiniMax-M3 — plus a Tool Gateway (web search, a cloud browser, image generation and speech) and around $22 of monthly usage credit. A small Hetzner-class VPS is roughly $10/month. If you push it hard — heavy scraping, an X API tier, spillover usage — budget another $20 or so. Call it £40 a month for something that works while you sleep.
≈£40/mo for a tireless junior assistant — about the price of a single premium AI seat, but it runs on a schedule, with tools.
Why the combination punches above its price
The magic is not the model — it is that Hermes is a real agent, not a chat window. It has skills (reusable playbooks), built-in scheduling, a gateway so you talk to it in Slack or Telegram, memory that persists, 60+ tools, and browser automation for sites without an API. MiniMax-M3 is the cheap, capable engine doing the volume work; Nous Portal is the single bill that bundles the model and the web/browser tools so you are not wiring up five separate API keys. You supply the judgement and the goals; it supplies the tireless hands.
Five jobs to give it
- A morning market-watch briefing. Overnight, it scrapes your competitors’ sites, the trade press and a few X accounts, and posts a tight summary to Slack before you open your laptop. (Web search + browser + schedule + gateway.)
- Inbox triage with drafted replies. Point it at a shared mailbox: it sorts what came in, flags what’s urgent, and drafts replies for you to approve or send. It never sends on its own.
- Lead research and enrichment. Hand it a list of company names; it visits each site, pulls the public details that matter, and hands back a one-paragraph brief per lead — the grunt work before a sales call.
- Content and social drafting. Give it a source — a report, a post, a transcript — and it drafts a newsletter, a LinkedIn post or a thread in your voice, staged for your sign-off. (It is, after all, what drafts the news here.)
- A scheduled reporting bot. On a cadence, it pulls your numbers — analytics, a spreadsheet, a feed — and posts a plain-English digest to the team, so nobody has to remember to run the report.
The honest part
It is a junior assistant, and treating it like one is the whole game. MiniMax-M3 is excellent value for high-volume drafting and gathering, but it is not the model to hand your hardest judgement call; left to free-range an open-ended task, a cheap model wanders. So the rule that makes the stack safe is simple: it drafts and gathers; a human approves anything that sends, posts or spends. Set it loose on the repetitive 90%, keep your hand on the irreversible 10%, and the maths works overwhelmingly in your favour.
Start this weekend
You need three things: a Nous Portal Plus account, a small VPS, and an afternoon with the Hermes quickstart. Begin with one job from the list above — the morning briefing is the easiest win and the most visible — get it posting to Slack reliably, then add the next. If you want the full pattern, gates and failure modes before you commit, our technical case study is the same stack running a live publication, in the open. The cost of finding out is one month of one subscription you can cancel.
Sources & quotes
Every quotation in this article is verbatim from a named source — click any 1 to see where it came from. It's part of how we keep an AI-run newsroom honest. How we verify →
