The Solopreneur’s AI Stack: How I Run 20 Agents for $96/Month

Last month, my 20 AI agents published 47 articles, optimized SEO on 200+ pages, generated 150 images, produced 3 podcast episodes, and sent 12 email campaigns. My total AI bill: $96.

Last updated: April 2026

Not $960. Not $9,600. Ninety-six dollars.

I am a solopreneur running seven websites, two YouTube channels, and an e-commerce tea shop from Hawaii. And the collection of ai tools for small business I have built over the past 18 months has completely changed what one person can accomplish. This is the behind-the-scenes story of how it works, what it costs, and how you can build something similar — even if you are starting from zero.

No hype. No gatekeeping. Just the real system, the real numbers, and the honest trade-offs.

The Backstory: From Burnout to Agents

Three years ago, I was doing everything by hand.

The visual below illustrates the complete roster of 20 specialized AI agents organized across content, SEO, and business operations.

Every blog post drafted in Google Docs. Every featured image sourced from stock photo sites or shot with my camera. Every product description written one at a time. Every meta description typed manually. Every internal link placed by scrolling through WordPress and hoping I remembered what I published last month.

I was running one website — All Day I Eat Like a Shark, a food blog focused on Japanese cooking and tea. One site, and I was already drowning. The content treadmill is real: you either keep publishing or your traffic decays. Google rewards consistency. But consistency at the pace content marketing demands is a full-time job by itself, and I had products to sell, videos to film, and a life to live.

So I did what most people do. I hired virtual assistants.

The results were… mixed. Good VAs are expensive. Affordable VAs need so much oversight that you barely save time. Training documentation goes stale the moment you write it. And the fundamental problem remained: I was still the bottleneck for every decision, every quality check, every "does this match our brand voice" judgment call.

Then I discovered AI agents.

Not ChatGPT — I had been using that for months to brainstorm and draft. I mean agents: persistent AI systems with specialized instructions, access to your tools, and the ability to execute multi-step workflows autonomously. The difference between asking ChatGPT to "write a blog post" and deploying an agent that monitors rankings, identifies gaps, drafts in your voice, optimizes for SEO, generates images, and publishes to WordPress — that is the distance between a calculator and a spreadsheet.

I built my first agent in early 2025. A content writer with instructions stored in a file called CLAUDE.md. Within two weeks, I knew this would reshape my entire business.

Eighteen months later, I run 20 agents across seven websites. They handle 80% of the operational work that used to consume my entire day. Total AI cost: $96 per month.

Here is exactly how it works.

The Three-Layer Stack

The single most important concept in my entire system is what I call the Three-Layer Stack. It is the framework that keeps costs low while keeping output quality high. Every task in my operation falls into one of three layers, and the layer determines what technology handles it.

If you want a deeper dive into this framework, I wrote an entire guide on it in my AI Content Creation Guide. But here is the essential version.

Layer 1: Zero-Cost Computation

Layer 1 tasks do not require any AI model. They are pure computation: scheduled scripts, data aggregation, API calls, file management, monitoring.

In my system, Layer 1 runs on my Mac through launchd (the macOS task scheduler) and a central orchestration script called data-pipeline.mjs. This single script coordinates dozens of automated workflows:

• Pulling Google Analytics and Search Console data every Monday at 6 AM
• Running site health audits across all seven websites monthly
• Syncing content calendars between Google Sheets and my project management tool
• Monitoring keyword rankings and flagging any drops greater than five positions
• Managing the publishing queue with daily cadence limits per site
• Backing up everything nightly

None of this needs intelligence. It needs reliability. A Node.js script running on a schedule does it perfectly.

Cost: $0. It runs on the Mac I already own.

Layer 2: Near-Zero Bulk LLM

Layer 2 handles tasks that need language understanding but not the best model available. I run Qwen3 32B (best for: cost-efficient bulk text generation and fine-tuning), an open-source LLM, locally on an NVIDIA RTX 5090 (best for: consumer-grade GPU computing with 32GB VRAM, supporting real-time model inference) through llama.cpp. Marginal cost per generation: electricity only, about $15-20 per month.

Layer 2 handles:

• First drafts of blog posts across all seven sites
• FAQ answer generation for schema markup (I batch-generated answers for 50 product pages in 20 minutes)
• Product description rewrites for SEO variations
• Sentiment analysis on product reviews
• Email newsletter summaries from blog content
• Alt text generation for thousands of images
• Meta description optimization

The same GPU runs ComfyUI (best for: node-based image generation workflow automation) with FLUX models for image generation, Whisper for transcription, and TTS for podcasts.

Cost: ~$18/month in electricity. No API fees. For about 70% of text generation tasks, Qwen3 32B is indistinguishable from premium models.

Layer 3: Premium Intelligence

Layer 3 is where I spend real money, and where it is worth every cent. These are tasks that genuinely require the best reasoning available: strategic planning, complex multi-step analysis, nuanced brand voice matching, quality evaluation, and creative direction.

I use Claude (best for: complex reasoning, multi-step agent orchestration, quality judgment on brand voice) from Anthropic as my Layer 3 brain — specifically Claude Code (best for: autonomous agent execution with file access and tool use), the CLI tool that turns Claude into an operating system for your business. It lets me build persistent agents with specialized instructions, tool access, and autonomous execution.

Layer 3 handles: strategic analysis (competitor gaps across 52 domains, quarterly planning), complex content that requires genuine reasoning, quality judgment before content goes live, agent orchestration, and structural editing.

Cost: ~$96/month in API fees. Some months $85, some months $110, but it averages to $96.

The key insight: by routing 70% of work to Layer 2 (free) and 20% to Layer 1 (free), only 10% of my workload hits the expensive model. That 10% is work that genuinely benefits from premium intelligence. According to ConvertKit’s 2024 research, 66% of creators used AI in 2023 (up from 34% in 2022), but few optimize their AI spend across cost-effective tiers — most creators default to premium-only models and overspend.

Meet the 20 Agents

Each agent is a specialized AI worker with its own instructions, tools, and domain expertise. Here is the full roster — the same team of ai tools for small business that runs my entire operation.

Content Production:

1. Content Writer — Researches topics, drafts WordPress-ready articles, follows site-specific templates and brand voice for all seven sites. Knows recipe formatting rules, internal linking patterns, and SEO requirements.

2. Content Refresher — Identifies stale content through ranking and traffic signals, merges new research with existing articles section-by-section (preserving what works), and republishes with updated dates.

3. Recipe Enhancer — Enriches recipe posts with culinary science, ingredient deep-dives, cultural context, pro tips, and common mistake warnings. Turns thin recipe pages into comprehensive resources.

4. Repurposer — Transforms blog posts into social media clips, podcast notes, email newsletter excerpts, and YouTube scripts. One piece of content becomes five distribution formats.

5. Newsletter Agent — Writes and schedules email campaigns through Omnisend (best for: e-commerce email marketing with segmentation and automation). Pulls from recent blog content, segments audiences, and manages drip sequences.

SEO and Technical:

6. SEO Optimizer — Audits pages for on-page SEO issues, fixes meta descriptions, optimizes heading hierarchy, and runs AEO (Answer Engine Optimization) passes for AI Overview and featured snippet targeting.

7. Schema Injector — Adds FAQ, HowTo, Article, and Recipe structured data to pages. Batch-processed 50 product pages with FAQ schema in one session.

8. Internal Linker — Scans the content corpus to find contextual internal linking opportunities. Adds 3-5 keyword-rich anchor links per post.

9. Alt Text Agent — Generates descriptive, keyword-aware alt text for images. Currently working through a backlog of 7,000+ images across all sites.

10. Site Auditor — Runs monthly health checks: broken links, missing images, thin content, orphan pages, schema errors. Produces actionable audit reports.

Research and Analysis:

11. Discovery Agent — Analyzes Google Analytics (best for: audience behavior tracking and content performance analysis) and Search Console (best for: search visibility, keyword ranking, and click-through rate data) data to find keyword opportunities, traffic trends, and content gaps. Feeds the content calendar.

12. Research Agent — Performs deep dives into topics, competitor strategies, and market trends. Produces research briefs that feed the content writer.

13. PAA Pipeline — Harvests "People Also Ask" questions from Google for target keywords, generates comprehensive answers via Qwen3, and injects them as FAQ sections with schema markup.

Business Operations:

14. Business Advisor — Weekly strategic briefings, revenue analysis, cross-site performance comparison. My Monday morning dashboard.

15. Growth Marketer — Analyzes advertising campaigns (Amazon Ads best for: product-level e-commerce advertising; Google Ads best for: search and display network reach), recommends budget shifts, flags wasted spend, and identifies scaling opportunities.

16. Ads Analyzer — Pulls campaign performance data, calculates ACoS/ROAS metrics, and generates optimization recommendations.

17. Listing Optimizer — Optimizes Amazon and Walmart product listings: titles, bullet points, descriptions, backend keywords. A/B test recommendations.

18. Review Bot — Monitors product reviews across platforms, performs sentiment analysis, drafts response templates, and alerts on negative reviews that need attention.

19. Cross-Sell Agent — Analyzes purchase patterns from WooCommerce (best for: self-hosted e-commerce with full customization) order data, identifies co-purchase opportunities, and suggests product bundles.

20. Cannibalization Resolver — Detects keyword cannibalization (multiple pages competing for the same term), recommends consolidation or differentiation strategies, and manages redirects.

Twenty agents. Each one specialized. Each one with clear instructions, defined boundaries, and quality gates. Together, they form a content and commerce operation that would require a team of 8-10 people to replicate manually.

The CLAUDE.md Framework: How Agents Get Their Instructions

Every agent in my system is defined by a single file: CLAUDE.md. This is the instruction manual that tells the agent who it is, what it does, how it makes decisions, and when to ask for help.

A well-written CLAUDE.md has six required sections:

1. Mission — What this agent does in one or two sentences. No ambiguity.
2. Tools — Which APIs, databases, scripts, and services the agent can access.
3. Decision Framework — How the agent prioritizes work and handles edge cases.
4. Output Format — Exactly what the deliverable looks like (WordPress HTML, JSON, Markdown, etc.).
5. Error Handling — What to do when things go wrong. Retry logic, fallback behavior, escalation rules.
6. Quality Gates — The standards that output must meet before it ships.

Here is a simplified example — the kind of CLAUDE.md that defines my SEO Optimizer agent:

# SEO Optimizer Agent

## Mission
Apply on-page SEO optimization to content across all sites.
Fix technical issues, optimize meta descriptions, add schema
markup, build internal links, and run AEO passes for
featured snippet and AI Overview targeting.

## Decision Framework
Check current rank before doing analysis:
- Rank #1 with snippet → Skip (already winning)
- Page 1 (#2-10) → Full SEO + AEO optimization
- Page 2 (#11-20) → SEO + lightweight AEO
- Page 3+ → SEO only (earn rank first)

## Quality Gates
- Meta descriptions: 120-155 characters, includes target keyword
- Internal links: 3-5 per post, contextual anchor text
- Schema: valid JSON-LD, tested against Google Rich Results
- Confidence threshold: >=0.85 auto-apply, <0.85 flag for review

That confidence threshold is critical. It is the mechanism that lets agents operate autonomously while keeping a human in the loop for judgment calls. When an agent is 85% or more confident that its work is correct, it executes automatically. Below that threshold, it flags the item for my review.

In practice, about 80% of agent output clears the confidence gate and ships without my involvement. The other 20% lands in my review queue in Mattermost (best for: self-hosted team communication and async collaboration), where I approve, reject, or modify it each morning. Without quality gates, you get AI slop. With them, you get a system you can trust.

The Hardware

Running 20 agents requires infrastructure. But it does not require expensive infrastructure. Here is the physical stack:

Mac (primary machine): Orchestration, scheduling, agent execution. Runs launchd cron jobs, data-pipeline.mjs, Claude Code agents, self-hosted n8n (best for: low-code workflow automation and API integration). Amortized ~$30/month.

NVIDIA RTX 5090 (remote GPU server): Bulk LLM inference, image generation, audio processing. Runs Qwen3 32B, ComfyUI + FLUX, WhisperX, TTS. Electricity ~$18/month, hardware amortized ~$25/month.

Hetzner VPS: WordPress (best for: flexible, extensible blogging and content management) hosting for all seven sites. MariaDB, OpenLiteSpeed, PHP 8.3, Redis on one instance. ~$25/month.

Proxmox Server: Grafana dashboards, Browserless (headless Chrome for scraping), nightly backups with offsite sync. ~$15/month.

Total hardware amortized cost: roughly $85-110/month.

You do not need all of this to start. I built up over 18 months. Your first agent needs just a laptop and a Claude API key.

The Real Cost Breakdown

Here is the honest accounting — the numbers I actually see in my bills and bank statements:

Category	Monthly Cost	What It Covers
Claude API	~$96	Layer 3 reasoning across all 20 agents
Qwen3 (own hardware)	~$18	Electricity for 5090 inference workload
ComfyUI (own hardware)	$0	Runs on same 5090, included above
n8n	$0	Self-hosted workflow automation (17 workflows)
WordPress hosting	~$25	Hetzner VPS for 7 websites
Proxmox server	~$15	Monitoring, automation, backups
Domain registrations	~$12	7 domains, amortized monthly
Total	~$166/month	Full operation: 7 sites, 20 agents, e-commerce

To replicate this with human workers — a part-time content writer, SEO specialist, VA, social media manager, email marketer, designer, and analyst — you would spend $8,000-12,000 per month conservatively.

I am not saying AI replaces all human judgment. I still make every strategic decision, film every video, taste every tea, and write the pieces that need my actual voice (like this one). But for the 80% of operational work that is mechanical — researching, drafting, optimizing, formatting, scheduling, monitoring, reporting — AI agents do it faster, more consistently, and at a fraction of the cost.

That is the real power of ai tools for small business operations. Not replacing yourself. Multiplying yourself.

What I Actually Do Each Day

Here is an honest weekday:

Morning (30-45 minutes):
I open Mattermost and review what the agents did overnight. My content refresher flagged three articles that need updated statistics. The SEO optimizer auto-fixed meta descriptions on 12 pages (confidence above 0.85, so it did not wait for approval). The review bot caught a negative product review and drafted a response. I approve two items, modify one response, and reject a draft that feels off-brand. Done.

Creative work (2-3 hours):
This is my actual job now. I film YouTube videos about Japanese cooking and tea. I work on book chapters. I record podcast episodes. I develop new tea blends and photograph products. This is the work only I can do — the work that should be my job, instead of writing meta descriptions.

Strategy (1 hour per week):
Monday morning, the business advisor agent generates a weekly briefing: traffic trends across all sites, revenue numbers, keyword ranking changes, content pipeline status. I read it, think about it, and make a few directional decisions. Should we lean into hojicha content this month? Is the garden site ready for a seasonal push? Where should the agents focus next?

I am the director, not the worker. The agents are my team. Some weeks I spend more time — quarterly planning, building a new agent, debugging a system. But the daily operational load has dropped from 8-10 hours to about 90 minutes. The rest goes to creative work, family, and the things that actually build long-term value.

How to Build Your First Agent

You do not need 20 agents to start. You need one.

Here is the path I recommend for any solopreneur or small business owner who wants to build their first AI agent:

Step 1: Pick One High-Volume Pain Point

Choose the task that eats the most hours of your week. For most content creators — whether you are looking into ai for content creators, e-commerce operators, or service businesses — it is one of these:

• Content drafting — writing blog posts, product descriptions, or social captions
• SEO optimization — fixing meta descriptions, adding schema, building internal links
• Email campaigns — writing and scheduling newsletters

Pick one. Just one.

Step 2: Set Up Claude Code

Claude Code is a command-line tool that lets you run Claude as a persistent agent with file access, tool use, and autonomous execution. Install it, get an API key, and run your first session.

If you have never used a terminal before, this will feel intimidating for about 15 minutes. Push through it. The learning curve is small and the payoff is enormous.

Step 3: Write Your CLAUDE.md

Create a CLAUDE.md file in your project directory with the six sections I described above: Mission, Tools, Decision Framework, Output Format, Error Handling, Quality Gates.

Start simple. You can always add complexity later. A 30-line CLAUDE.md that clearly defines what the agent does is better than a 300-line document that tries to cover everything.

Step 4: Test With Human Review on Everything

For the first two weeks, set your confidence threshold to 1.0 — nothing auto-publishes, everything goes through your review. Read every output. Refine the CLAUDE.md based on what you observe. This calibration period is where the agent goes from generic to genuinely useful.

Step 5: Gradually Increase Autonomy

Once you trust the output, lower the confidence threshold. Let the agent auto-execute on tasks where it consistently performs well. Keep human review on anything that drifts. Over time, most routine work clears the gate automatically, and you start thinking about agent number two.

Step 6: Add Layers

When your first agent is reliable, add a local LLM (Ollama or a budget API) for bulk tasks, add scheduling with cron or Task Scheduler, and build agent number two. The system grows organically, one agent at a time, each one reclaiming a few hours of your week.

The Honest Trade-Offs

I would be lying if I said this was all upside. Here are the real trade-offs:

Setup time is real. I spent 18 months building this system. The first agent took a weekend. The full operation took hundreds of hours. This is an investment that pays compounding dividends — like planting a garden.

Technical skill helps. I am comfortable with the command line and Node.js. If you are not, the curve is steeper — but Claude Code is designed to be approachable, and the AI can help you build the very system it runs on.

Quality requires vigilance. AI agents will confidently produce mediocre work if you let them. The quality gates and confidence thresholds are not optional. They are what separate a useful ai automation business system from a content mill that publishes garbage.

Hardware is an upfront cost. The RTX 5090 was $2,000. You can start without it — just a laptop and a Claude API key — but the full three-layer stack requires some hardware investment that pays for itself within two to three months.

The Bigger Picture

Here is what I keep coming back to, and the reason I am sharing all of this publicly.

The conversation around ai tools for entrepreneurs usually splits into two camps: AI will replace everyone and you should panic, or AI is overhyped and you should ignore it. Both are wrong.

AI agents are not replacing me. They are freeing me. The mechanical work — SEO, meta descriptions, schema, alt text, email formatting — consumed my creative capacity. Now it happens in the background while I film a video about brewing hojicha or write a chapter of my next book.

That is the promise of AI for solopreneurs. Not "do less." Not "get rich while you sleep." But: spend your time on the work that only you can do, and let systems handle the rest.

It costs me $96 a month in AI and about $166 all-in. Twenty agents across 7 websites, getting better every month.

If you are a solopreneur drowning in operational work, this is the exit. Not another VA. Not another course. A system — built once, improved continuously, compounding in value every single day.

Start with one agent. Write your CLAUDE.md. Test it for two weeks. And watch what happens when you stop being the bottleneck in your own business.

Want the exact CLAUDE.md configurations I use for all 20 agents?

I have packaged my complete CLAUDE.md Template Library — all 20 agent configurations, the Three-Layer Stack setup guide, and the confidence-gating framework — into a free download.

Download the CLAUDE.md Template Library (20 Agent Configurations) →

It is the same system powering the operation you just read about. No email opt-in wall. No upsell. Just the templates.

---

Frequently Asked Questions

What is Claude Code, and how is it different from ChatGPT?

Claude Code is a command-line interface that turns Claude (Anthropic's AI model) into a persistent agent with file access, tool integration, and autonomous execution capabilities. Unlike ChatGPT, which is primarily a conversational interface, Claude Code lets you deploy agents that can execute multi-step workflows, maintain state, and make decisions autonomously. The agents can read and write files, call APIs, and operate on schedules without your intervention between tasks.

How much does it cost to start building AI agents if I don't have the hardware?

You can start with just a laptop and a Claude API key — total cost: $0 upfront, then Claude API usage (typically $10-30/month for a single agent). The hardware stack I describe (GPU, additional servers) is optional and comes later as you scale. Most solopreneurs start with one Layer 3 agent (Claude) running on their existing machine and add local hardware only when the API costs justify the investment.

How long did it take you to build this 20-agent system?

18 months total, but not all at once. The first agent took a weekend. By month 6, I had 5 agents. By month 12, I had 15. By month 18, I had optimized the stack to 20 specialized agents. You don't need 18 months to see ROI — most solopreneurs report meaningful time savings within 2-4 weeks of deploying their first agent. My timeline reflects continuous optimization and adding complexity, not the minimum viable system.

What programming skills do I need to build agents?

You should be comfortable with basic terminal commands and able to write/edit text files (like CLAUDE.md). You don't need to be a software engineer. Claude Code is designed to be approachable for non-technical users, and Claude itself can help you debug issues and improve your agent instructions. If you've used a Mac or Linux terminal before, you have enough skills to start. If not, expect a 1-2 week learning curve.

What happens if an agent produces poor-quality output?

That's what confidence thresholds and quality gates are for. Every agent in my system has a confidence score. If it's below 85%, the output goes to a human review queue (my Mattermost channel) instead of publishing automatically. I review it each morning, approve it, modify it, or reject it. About 80% of agent output clears the bar automatically; the other 20% gets human judgment. This prevents publishing garbage while keeping most work autonomous.

Pat Tokuyama is the founder of All Day I Eat Like a Shark and Digital Garden Profit. He runs seven websites, two YouTube channels, and a Japanese tea and kitchenware business from Hawaii, powered by 20 AI agents and a stubbornly anti-hustle philosophy. Connect with him on YouTube or at digitalgardenprofit.com.

---