"$0 in LLM costs." That's what a Taiwanese solo developer posted on Hacker News — and it caught a lot of attention. He's running 4 AI agents on nothing but Gemini 2.5 Flash's free tier, automating content creation, sales lead generation, security scanning, and operations management. 27 social accounts. 3.3 million views. Monthly infrastructure cost: $5.
What Is It?
The developer goes by Bo Liu, and he runs a one-person tech agency in Taiwan. Using OpenClaw (an open-source AI agent platform) on WSL2, he's built 4 agents — each with a clearly defined job:
| Agent | Role | Key Tasks |
|---|---|---|
| Content Agent | Social media post creation | 8 posts/day with self-review quality gates |
| Sales Agent | Lead generation | Security scanning to find leads, customer data sync |
| Security Agent | Security vulnerability scanning | Prospect site analysis, report generation |
| Operations Agent | Day-to-day operations management | Endpoint monitoring, data sync |
Here's the thing — what makes this setup special is the core design principle: "Use the LLM only for creative tasks — handle all data collection over HTTP." RSS feeds, Hacker News, web scraping — none of that research pipeline touches a single LLM token. It all runs as pure HTTP calls using tools like Jina Reader.
What Changes?
You've probably heard plenty of "automate your business with AI agents" pitches. Here's what makes this one different.
First, the LLM costs really are $0. Most AI automation case studies that claim to be "cheap" are quietly hiding $50–$200+ in monthly API bills. Bo Liu is using just 105 of Gemini 2.5 Flash's 1,500 free daily requests — which means 93% of the daily limit is still untouched.
Second, this is concrete proof that the "one-person company" model actually works. The One-Person Company (OPC) concept Bo Liu writes about in his Medium post isn't abstract futurism — it's a system running right now. Agents fill the roles of content manager, sales rep, security engineer, and operations manager.
Why Gemini's free tier?
Gemini 2.5 Flash is free on Google AI Studio, with limits of 1,500 requests/day and 30 RPM. The key strategy is simple: "Pack all context into one request, get one complete response." Instead of multi-turn conversations, Bo Liu pre-processes data into local Markdown files (zero token cost) and feeds everything into a single LLM call.
| Comparison | Typical AI Automation | Gemini Free Tier Stack |
|---|---|---|
| LLM Cost | $50–$200+/month | $0 |
| Infrastructure Cost | $20–$100/month | ~$5 (Vercel + Firebase free tiers) |
| API Call Strategy | Multi-turn conversations, chaining | All context in a single call |
| Data Collection | LLM-based parsing | Pure HTTP (zero token cost) |
| Scheduling | External cron services | 25 systemd timers |
Bo Liu also shared a mistake he made — he accidentally used an API key tied to his GCP billing account and ended up with a $127 bill. He'd confused his AI Studio key with a GCP project key. That's the hidden trap with the free tier.
Heads Up: Free Tier Risks
Gemini's free tier only applies to keys created in Google AI Studio. Keys generated from the GCP console will be billed — even for the exact same model. Mix them up and you could see hundreds of dollars disappear overnight. Also note that free tier usage may have restrictions on commercial use, so check the terms of service before you go all-in.
Getting Started
- Get your Gemini API key
Head to Google AI Studio (ai.google.dev) and create a free API key. It must be from AI Studio — not the GCP console. You get 1,500 free requests per day and access to the Gemini 2.5 Flash model. - Separate your data pipeline from the LLM
This is the core design principle. For data collection tasks — RSS feeds, web scraping, API calls — don't involve the LLM at all. Use pure HTTP with tools like Python's requests, feedparser, or Jina Reader. Save LLM tokens exclusively for reasoning and generation. - Apply the "one request = one result" pattern
This is how you work around the free tier's RPM limits. Pre-collect your data, organize it into local Markdown files, then inject all that context into a single LLM call and get your result back in one response. Multi-turn conversations burn both tokens and time. - Set up OpenClaw or a similar platform
OpenClaw is an open-source AI agent platform. Install it on WSL2, Docker, or a server, connect your Gemini API key, and start building agents. That said, as Bo Liu himself noted, getting it to production level requires significant customization. - Start small, then scale
Don't try to build all 4 agents at once. Start with a content agent — "collect trending topics from RSS → draft with Gemini → self-review → post to social account." Once that pipeline is stable, layer in the next agent.
