A case study went around recently: a lawyer had wired up 66 AI agents on a Mac mini for his own firm — every one running locally, nothing touching a cloud API — and was looking for a commercial partner before releasing it as open source.
I read that and realized I had been building the same shape of thing for my own small business for the last year. Eight ambient agents, all running on Apple Silicon, none of them touching cloud APIs. I had not productized any of them. I had also not gotten paid for any of them.
This post is about what I'm doing about that.
The asset base
I shipped a repo called claude-code-local. It's an MLX server that wraps three local language models (Gemma 4 31B, Llama 3.3 70B, Qwen 3.5 122B MoE) behind an OpenAI-compatible API. The setup script picks the right model for your hardware, downloads it, and puts a launcher on your Desktop. Three commands and you're running a 31-billion-parameter language model on your MacBook.
It has 2,689 stars and 516 forks as of this writing. License is MIT. The whole thing is at github.com/nicedreamzapp/claude-code-local.
That's the engine. What I had not built was the funnel.
The gap that ate me for a month
Stars are people who think "I would love this." Forks are people who started doing the work. Neither converts to revenue.
Reading that case study, what jumped out was: a lawyer with the technical chops to wire 66 agents could not productize the result. He has the buyer relationships (he is the buyer); I have the go-to-market background (I've sold consumer hardware direct to customers for years). The intersection nobody is shipping is "Mac mini with this stack pre-installed, delivered to your law firm."
So I sat down and mapped what was actually missing.
The market gap, in numbers
I researched the on-device AI market for May 2026. Here is what I found:
- Free OSS (Ollama, LM Studio, Jan, Atomic Bot, my own repo): saturated, no money flowing
- Paid Mac App Store apps (Private LLM, Enclave AI, Local LLM): $10-30 one-time, real revenue for bootstrapped 2-person teams
- Cloud SaaS with zero-data-retention (Spellbook, CoCounsel): $69-$149/mo per seat, cloud-not-local
- Enterprise legal AI (Harvey): $1,200+/seat/month
- AI-native law firms (Manifest, Avantia, General Legal): not selling tools, they ARE the firm
Two things stand out:
The gap between "free OSS" and "$1,200/seat SaaS" is not being filled by subscription products. Private LLM's App Store reviews explicitly call out "no subscription" as the reason buyers picked them. The privacy buyer rejects recurring billing for privacy software. This is not opinion; it is in their reviews.
The Mac mini hardware bundle for privileged work is being built privately but not productized. Nobody is shipping it as a SKU.
That is the gap. That is what I am putting a product against.
The ladder
I'm skipping the standard SaaS playbook because the market is telling me to. The privacy buyer rejects recurring billing — so the spine is one-time purchases, not subscriptions:
| Rung | Price | Audience |
|---|---|---|
| Free repo | $0 | Existing OSS audience, top of funnel |
| AirGap Box | $2,995 (Base) / $3,995 (Pro) | Small firms wanting sovereignty without DIY |
| Foundation consulting | scoped | Firms ready for deeper, white-glove deployment |
The free repo gets you curious. The Box converts the firms who would rather pay $3k than learn MLX. Foundation is for the firms that want it installed, tuned, and documented for their compliance counsel.
The three core agents
Three agents that drop on top of any OpenAI-compatible local LLM server (the free claude-code-local repo by default, but they work with Ollama and others too) — and ship pre-installed on the Box:
Folder Watcher — drop a PDF, text, or Markdown file into ~/AirGap-Inbox/. Within 30 seconds, a structured Markdown summary appears in _summaries/. Uses macOS's textutil and mdimport for PDFs and docx without external dependencies.
Daily Briefing — every morning at 7:00 a LaunchAgent reads the folders you list in a config file and writes a one-page digest of what changed in the last 24 hours.
Local Q&A — a CLI command
airgap ask "your question"that answers from a single document you point it at, fully offline. Folder-wide indexing and citations across many files is on the roadmap — not something I'd oversell today.
The whole set is ~300 lines of Python. The installer is a .command file you double-click. Total install time: under 60 seconds on the right hardware.
I deliberately built them with zero authentication required. No IMAP credentials, no OAuth dance, no API keys — the install works for everyone on day one. The agents that need credentials (email drafting, Reddit lurking, calendar parsing) run on my own machines and ship in the AirGap Box, where there's a setup call to wire them up properly.
What's in the AirGap Box
The same stack on a Mac mini that arrives at your office. Pre-installed: claude-code-local MLX server, the three core agents, two more that need credentials (email drafter, prompt library), a default-blocked firewall, and a printed compliance memo template.
Base ($2,995): Mac mini M4 16GB, Gemma 4 31B preloaded.
Pro ($3,995): Mac mini M4 Pro 24GB, Llama 3.3 70B preloaded.
Both include a 90-minute Zoom setup call and 30 days of email support. After that, you have a working private-AI workstation. We do not phone home. We do not see your data.
COGS on the base unit is around $940. Margin around 69%. Cash-flow safe because Stripe charges at purchase; hardware ordered after.
Validation before scale
I am not ordering inventory on speculation. The plan is:
- Day 0: launch publicly, with the Box gated by a waitlist (no checkout yet)
- Day 7: first review — Box waitlist signups
- Day 14: hard gate on Box — 20+ verified emails or we revisit positioning
- Day 30: full P&L review, decide whether to scale
If the Box waitlist hits 20 in 14 days, I order the first 3 Mac minis. If it hits 5, I have not validated the rung; I revisit positioning before spending hardware money.
This is the part the grifter posts skip. "I made $14,200/month in 72 hours" is not a thing that happens to honest businesses. What happens to honest businesses is "I opened a waitlist, watched signups for two weeks, decided whether to order inventory based on actual demand, and reported the real numbers."
Honest year-1 range
| Month | Pessimistic | Realistic | Stretch |
|---|---|---|---|
| 1 | $0 | $400 | $2,000 |
| 3 | $200 | $1,500 | $6,000 |
| 6 | $800 | $3,500 | $12,000 |
| 12 | $1,500 | $6,000-8,000 | $20,000+ |
To hit the grifter's $14k/month claim, every rung needs to perform near the top of its range AND consulting needs to land. That happens in months 12-18 with discipline, not in 72 hours.
What I will publish honestly
- Weekly waitlist + signup numbers (real, not vanity)
- The first refund (when it happens) and why
- The first Box install case study (with the firm's permission)
- The Box waitlist → order conversion, as it happens
If you want to see whether this business works, follow along on github.com/nicedreamzapp/claude-code-local and the AirGap landing pages. I'll post the numbers as they happen.
Links
- Free repo: github.com/nicedreamzapp/claude-code-local
- AirGap Box waitlist: nicedreamzwholesale.com/airgap-box
- Demo (NDA review on a laptop with Wi-Fi physically off, lsof on screen): youtube.com/watch?v=V_J1LpNGwmY
Why I'm posting this
Because the next person to read a case study like that and think "I want this" deserves to find a productized version, not another Hacker News thread about MLX setup. And because every honest version of this story I publish is also a receipt — for me, that I built something real, and for the next builder, that the playbook works.
If you have feedback on the pricing, the positioning, or the agents, leave a comment. I read everything.
I'm building AirGap — pre-configured local AI on a Mac mini for firms handling private work. Join the AirGap Box waitlist, grab the free open-source stack, or see AirGap consulting for compliance-sensitive firms (law, medical, finance).
Top comments (0)