March 31, 2026 · ~7 min read
Yesterday I had a personal agent. Today I have a business prototype.
What started as "what if people could have their own Harvie?" ended with a working bot, a real-cost benchmark and the first Apify scrapers actually making money (instead of losing it).
The morning briefing already runs on its own
At 8 in the morning, before I even opened my eyes, Harvie had already:
- Synced my HRV and COROS data
- Read tweets from @ErickSky, @steipete, @carlosvillu, @openclaw and @PerSolana
- Downloaded the BOE RSS feed and filtered relevant provisions
- Prepared my workout for the day (strength at home: core, glutes, ankle)
- Summarized the Noticias Ilustradas video
All via Telegram. I just had to read it.
The million-dollar question: how much does an agent cost?
My agent runs on Claude Opus 4.6 — Anthropic's most powerful model. That costs about €275/month in API alone. Works great, but doesn't scale.
So we ran a serious benchmark: 8 tests (chat, tool calling, planning, opinion, technical analysis, security) comparing:
- Claude Opus 4.6 — ~€275/month, perfect tool calling, slow
- MiniMax M2.7 — ~€15/month, tool calling 2/2, normal speed
- DeepSeek V4 — ~€12/month, tool calling 1/2, fast
MiniMax wins. At $0.30 per million input tokens and $1.20 output, it's 18 times cheaper than Opus. And the key part: when you tell it "search for flights to Tokyo and notify my wife on Telegram", it does both things. DeepSeek only does one.
Casandra Lite: the MVP
With that data, we built Casandra Lite — a multi-tenant Telegram bot running MiniMax M2.7 with native tool calling.
What it has:
- Invitation code system (CASANDRA-XXXXX)
- Conversational onboarding
- Chat with MiniMax as the brain
- Real tools: web search (Brave), weather (wttr.in), reminders, URL reading
- Telegram admin panel (/admin)
- Real-time supervision of all conversations
- Tier limits (free: 20 msgs/day, pro: unlimited)
- systemd service that doesn't crash
What it cost to build: one day. Four Python files, SQLite and a systemd service.
What it costs to run: ~€15/month per user in API + a shared VPS.
The business model is simple: invitation codes, Bizum for payment, €29/month per user. No website, no Stripe, no friction. The first users will be people close to us who try it and give us feedback.
Apify: from losing money to making it
I have 5 scrapers published on Apify. The TikTok one was losing $1.33 per run — $0.32 in revenue against $1.65 in cost.
The problem: 4GB of RAM, 1-hour timeouts, networkidle burning 15 seconds per page, and duplicate data in the output.
The fixes:
- RAM from 4GB to 2GB across all 5 actors
- Timeout from 1 hour to 10 minutes
networkidle→domcontentloaded(saves ~10s per page)- Removed the
_rawfield that was duplicating all data - Updated pricing to cover costs + 12% margin
Now the TikTok scraper charges $0.35/GB start + $0.03 per result. With 2GB of RAM, each run generates revenue instead of losses.
What I learned today
-
Cheap models are surprisingly good for 90% of personal assistant tasks. You don't need Opus to answer questions, search the web or create reminders.
-
Native tool calling is key. Simulating tools with text prompts is fragile. Letting the model call functions directly works much better.
-
Cost per user determines the entire business model. At €275/month there's no business. At €15/month, with a price of €29/month, there's room to grow.
-
Scrapers need maintenance. Publishing an actor on Apify and forgetting about it is a recipe for losing money. You need to monitor costs and adjust prices.
What's next
- Test Casandra Lite with 3-5 real people
- Add more tools (email, calendar)
- Improve the YouTube and LinkedIn scrapers
- Explore the Apify plugin for OpenClaw (access to thousands of Actors from chat)
- And keep documenting everything here
If you're interested in trying a personal agent or building your own, I'm on Twitter as @Gantit. DMs open.
— I, Johnny — configured agent: Harvie. If you automate what doesn't matter, you're left with time for what does.