7 AI Agents I Run on a $600 Mini-PC for $18/mo

By Lazar Milicevic · Published September 25, 2026 · 9 min read

Most solo founders I talk to pay $300-$500/month for a stack of SaaS tools that overlap by 60%. Zapier for plumbing, Lindy or Relay for agents, a CRM for follow-ups, an invoicing tool, a PR bot, a newsletter aggregator. I run all of that from a mini-PC on a shelf in my apartment for $18/month in Claude API calls. Here's the exact stack, the seven agents that ride on top of it, and the numbers each one produces.

The box: $600 once, then $18/month forever

The hardware is a small-form-factor mini-PC I paid about $600 for — 32GB RAM, 1TB NVMe, a modern low-power CPU. Nothing exotic. It runs Windows on the metal and Ubuntu inside WSL2, because I want the Linux tooling without dual-booting my daily driver. On top of that:

• n8n (self-hosted, Docker) — the orchestrator. Every agent is a workflow triggered on a cron, a webhook, or a Telegram command.
• Claude API — the reasoning layer. Every agent that needs to classify, draft, or summarize hits Anthropic's API.
• Telegram Bot API — the interface. One chat app on my phone talks to all seven agents. No dashboards, no browser tabs.

The Docker compose that starts the whole thing is 40 lines:

services:
  n8n:
    image: n8nio/n8n:latest
    restart: unless-stopped
    ports:
      - "5678:5678"
    environment:
      - N8N_HOST=localhost
      - N8N_PROTOCOL=http
      - GENERIC_TIMEZONE=America/New_York
      - N8N_ENCRYPTION_KEY=${N8N_KEY}
    volumes:
      - ./n8n_data:/home/node/.n8n
      - ./files:/files

That's the entire infrastructure. n8n is free when you host it yourself. Telegram is free. The mini-PC pulls maybe 15W idle, so electricity is a rounding error. The only recurring bill is Claude API usage, which for the last three months has averaged $17.80/mo across all seven agents combined.

Why so cheap? Because I use Claude Haiku for classification and enrichment (cheap, fast), and only escalate to Sonnet for anything that needs actual reasoning — draft emails, PR reviews, weekly summaries. About 85% of my calls are Haiku.

Agent 1: Inbox Chief — 120 emails in, 12 out

Every morning at 7:00 AM, n8n fires a workflow that pulls unread Gmail via the Google API, hands each message to Claude with a classification prompt, and drops the result into four buckets: client, lead, invoice-related, noise. Noise gets auto-archived. Client messages get a suggested reply drafted underneath. The whole thing arrives as one Telegram message before I've had coffee.

Real numbers from last month: average 118 emails processed per day, I read 11 of them. The classification prompt is 200 tokens, the average email is 400 tokens in, 150 tokens summary out. At Haiku pricing that's roughly $0.0004 per email, or about $1.40/month for this agent alone.

The prompt structure I use:

You are triaging email for a solo operator.
Return JSON: {bucket, priority, one_line_summary, suggested_reply?}

Buckets:
- client: existing paying customer
- lead: prospect or inbound inquiry
- invoice: payment, receipt, or financial
- noise: newsletters, notifications, cold outreach

Draft suggested_reply ONLY for bucket=client with priority=high.

The trick that made this actually usable: I keep a whitelist.json of email addresses that always bypass classification (my accountant, key clients, my mom). Never trust an LLM with 100% of your triage.

Agent 2: Lead Hunter — 40 pre-written drafts/week, zero sends

This one runs three times a day. It hits a list of sources I've whitelisted — industry directories, specific subreddits, job boards where my ICP posts, a few LinkedIn feeds via a scraping API — pulls new entries, enriches each one with a quick web lookup (company size, tech stack, recent news), and drafts a personalized first-touch email.

It does not send. That's deliberate. Drafts land in a Telegram queue with inline buttons: Approve / Kill / Edit. Approve routes the draft through my normal Gmail via SMTP so it inherits my sender reputation and doesn't trip spam filters. Kill discards it. Edit opens the draft in a follow-up message I can reply to.

Over 12 weeks: 487 drafts generated, 312 approved and sent, 41 replies, 9 discovery calls booked. That's a 13% reply rate on cold outbound, which is well above industry norms — because every draft references something specific from the enrichment step ("saw you just hired two BDRs" beats "hope this finds you well").

Agent 3: Invoicing — 15 minutes to 15 seconds

I send one Telegram message: /invoice AcmeCorp $4200 "Q1 automation build". The agent generates the invoice PDF, applies the correct tax treatment for my jurisdiction, emails the PDF to the client's billing contact, and files a copy in the right Google Drive folder named by client and quarter.

This is the agent I always highlight when someone asks "what can't SaaS do for me?" Generic invoicing tools don't understand the quirks of every country's VAT rules, reverse-charge mechanics, or the specific layout your accountant needs to import cleanly. When you own the code, you fix it in an afternoon. When you rent the SaaS, you file a feature request and wait forever.

The layout is a simple HTML template rendered to PDF with weasyprint, tax logic is a 60-line Python module, and the whole workflow is one n8n node calling a local script. Total build time: about 6 hours. Total time saved: ~1 hour per week, forever.

Agent 4: Follow-Up Bot — days-to-payment 31 → 19

The Follow-Up Bot watches the invoice folder that Agent 3 writes to. Each invoice has a due date. Cron runs daily and checks:

• Day 14 unpaid: friendly nudge, cheerful tone, "just floating this back up"
• Day 21 unpaid: firmer, references the original due date explicitly
• Day 30 unpaid: pings me on Telegram, no auto-email — I handle it personally

Three months of running it, my average days-to-payment dropped from 31 to 19. That's a 39% reduction in collection time with zero mental effort on my part. For a solo business where cash flow is oxygen, that alone is worth more than the entire stack.

The reason it works isn't the AI — it's the discipline of always sending on day 14. Humans forget. Cron doesn't.

Agent 5: PR Reviewer — first-pass review in 90 seconds

If you ship code, you know the loop: PR opens, you're mid-flow on something else, it sits for two days, momentum dies. This agent subscribes to my GitHub webhooks. Every new PR triggers a workflow that:

1. Fetches the diff via the GitHub API
2. Loads a conventions.md file from the repo root
3. Asks Claude Sonnet: "Review this diff against these conventions. Flag missing tests, error handling gaps, unused imports, and inconsistencies with the existing codebase."
4. Posts the review as a PR comment via the GitHub API

The conventions.md file is where the leverage is. Mine is 80 lines: naming conventions, error handling patterns, test coverage expectations, banned imports. The review isn't smarter than me — it's just faster and never forgets. When I open the PR, half the boring work is done and I'm making judgment calls, not spotting typos.

Agent 6: Content Radar — 45 min doom-scroll → 2 min skim

Every evening at 6:00 PM, one Telegram message: the ten things worth knowing today, one line each, with a link. Sources: Hacker News top 30, a curated list of ~40 X accounts, and 8 YouTube channels via their RSS feeds.

The pipeline:

fetch sources → dedupe → score by relevance to my interests
→ pick top 10 → summarize each in one line → Telegram

Scoring uses a simple prompt: "Rate 0-10 how relevant this is to a solo founder building AI automation for SMBs." Anything scoring 7+ makes the cut. The agent also has a watchlist.txt of competitor names and tools I'm tracking — anything matching gets flagged with a 🔔 so I never miss a launch.

I used to spend 40+ minutes a day scrolling to feel current. Now it's a 2-minute skim while the espresso pulls.

Agent 7: Weekly Ops Report — the meta agent

Every Monday at 8:00 AM, this one queries the logs of the other six and builds a single Telegram message:

Week of Nov 17 Emails triaged: 743 (12% required attention) Leads drafted: 131 · Approved: 84 · Replies: 11 Invoices sent: 22 · Overdue flagged: 3 PRs reviewed: 7 Links surfaced: 70 · Watchlist hits: 4

I read it in 30 seconds and I know exactly where the business stands before my first meeting. This one costs about $0.20/month to run — one Sonnet call per week.

The real cost comparison

Here's what the equivalent commercial stack runs, roughly:

My setup: $600 hardware once, $18/month recurring. Payback vs. the low end of that range is about 2.5 months. Vs. the high end, under 6 weeks.

That's not counting the compounding cost: every SaaS you add pulls you deeper into their pricing tier ladder. Every self-hosted agent I add costs me maybe $2-4/month more in API calls, and that's it.

What breaks and what I'd do differently

Honest about failures: n8n has crashed on me twice in a year, both times because I let disk fill up with old execution logs. Fix: a nightly cron that prunes anything older than 14 days. The Lead Hunter has drafted embarrassing emails twice — once addressing someone by their company name, once referencing a "recent" article from 2019. Fix: I added a sanity-check step that flags drafts where the enrichment data is thin or contradictory.

If I were starting today, I'd skip WSL and run Ubuntu bare-metal on the mini-PC. WSL works but adds a networking layer you'll fight eventually. And I'd put the whole thing behind Tailscale from day one so I can access n8n from anywhere without exposing ports.

Where bizflowai.io fits in

This is exactly the kind of stack I build for clients week to week at bizflowai.io — custom agent workflows on infrastructure they own, wired into whatever they already use (Gmail, HubSpot, QuickBooks, GitHub, Slack). The pattern is always the same: one orchestrator, one LLM provider, one messaging interface, and a handful of well-scoped agents that each do one job well. No 12-app SaaS zoo, no per-seat pricing that punishes growth.

Start with one agent, not seven

If you're reading this and thinking about building your own version: don't try to ship all seven at once. Pick the single most painful workflow in your week — for most solo operators that's inbox triage — and build only that. Get it stable for two weeks. Then add the next one. That's how I built this stack, and it's the only way I've seen anyone actually finish.

Want more like this?

I publish practical AI automation, GenAI engineering, and faceless content workflows on YouTube every week.

Subscribe to bizflowai.io on YouTube — never miss a new tutorial.

Planning an AI automation project or need a second opinion on your architecture?

Connect with me on LinkedIn — Lazar Milicevic, GenAI Engineer & bizflowai.io Founder.

Visit bizflowai.io for our services, case studies, and AI consulting.