In March 2025, Columbia University suspended a student named Chungin “Roy” Lee. His crime: building Interview Coder, a Chrome extension that feeds AI-generated solutions to LeetCode-style problems during live technical interviews — through an overlay the screen recorder doesn’t capture.

Eighteen months later, that “cheating tool” is a company called Cluely, with $5.3 million in seed funding, a $15 million Series A led by Andreessen Horowitz, $20.3 million total raised, and 70,000 signups in its first week after launch in April 2025. Lee claimed $7 million in annual recurring revenue on stage. He later admitted the real number was $5.2 million — still real revenue, from a product whose original pitch video was literally titled “cheat on everything.”

You can argue about the ethics all day. What you can’t argue about is the market signal: people will pay monthly for an AI that watches their screen and tells them what to say, invisibly. And the entire technical stack needed to build one runs on free tiers you can sign up for tonight.

This article is the playbook: what an invisible overlay actually is, the verified free-tier limits that power it, the architecture, and the part nobody writes about honestly — how you collect money for a product that Stripe and the platforms would rather didn’t exist.

What an “invisible overlay” actually is

Strip away the hype and Cluely, Interview Coder, and every clone in between are the same four components:

  1. A frameless, always-on-top window that floats over everything. Transparent background, no taskbar entry, global hotkey to summon it.
  2. Screen capture — a region screenshot or a rolling buffer of the active window, sent to a vision-capable model.
  3. A model call — “here’s the question on screen, here’s my resume context, answer in under 80 words.”
  4. The invisibility trick. On Windows, the SetWindowDisplayAffinity API with the WDA_EXCLUDEFROMCAPTURE flag tells the OS compositor to skip your window when any other process captures the screen. Zoom, Meet, and most recorders never see it. On macOS the equivalents live in private CoreGraphics APIs and virtual display tricks; on Linux, a compositor layer does it trivially.

That fourth component is public OS documentation, not a hack. Screen-recording exemptions exist for legitimate reasons — password managers, private windows. The overlay product category just points that machinery at interviews, sales calls, exams, and meetings.

The result: the user sees answers floating over their interviewer’s face. The interviewer sees a normal candidate looking at their webcam.

The demand is measurable, not vibes

If you think this is a niche of lazy CS students, look at the recruiting side. HeroHunt.ai, a recruiting-tech company, analyzed 19,368 AI-led interviews run between July 2025 and January 2026 and flagged 38.5% of cases for suspected AI assistance. Recruiters are so spooked they’re buying detection tooling — which means candidates are buying countermeasures, which means better detection, forever. It’s an arms race with two paying sides.

Even the giants validate the format and fumble it. OpenAI shipped ChatGPT Atlas, a browser with an agent in the sidebar, in October 2025 — then shut it down in July 2026, folding the idea into a Chrome extension. The concept (AI hovering next to your work, reacting in real time) survives; the packaging keeps getting rebuilt. Independent builders don’t need to ask permission to rebuild it cheaper.

The free-tier stack, with real numbers

Here’s what powers a working overlay assistant for $0/month, verified against provider docs and rate-limit trackers as of October 2026:

Gemini CLI — 1,000 requests a day, free. Sign in with a personal Google account and Gemini CLI’s free tier allows 1,000 requests per day (250/day if you authenticate with an unpaid API key instead, Flash only). A CLI agent sounds like a dev tool, but nothing stops your overlay backend from shelling into it or reusing the same OAuth’d quota through the API. One thousand answered questions a day is more than any human asks during interviews and meetings combined. One Google account = one seat of capacity, for nothing.

Cloudflare Workers AI — 10,000 Neurons a day, free. Cloudflare’s inference platform gives every account 10,000 Neurons daily at no charge, on both the free and paid Workers plans, with overage at $0.011 per 1,000 Neurons. That’s enough for a steady stream of small-model completions or captioning — and it’s a second independent quota, from a second provider, under a second account. It resets daily at 00:00 UTC.

Anthropic web search — $10 per 1,000 searches. When the overlay needs to look something up live (the “what is this company’s pricing?” moment on a sales call), Anthropic’s web search tool costs $10 per 1,000 searches plus token costs, per their official pricing docs. At even 10 lookups per active user per day, that’s one cent per user-day — billable against a $20/month subscription a hundred times over.

Do the unit economics on one paying user at $25/month:

  • 300 overlay queries/day × 30 days = 9,000 requests. Two-thirds absorbed by the Gemini CLI free 1,000/day; the rest spills into Workers AI neurons or a paid Flash key at fractions of a cent.
  • Live lookups: ~300/month ≈ $0.30 at Anthropic’s rate.
  • Infrastructure: a $5 VPS runs the routing layer for dozens of users.

Gross margin sits above 95% before you’ve written the billing page. This is why VCs funded it: the cost of intelligence has collapsed to the point where the product is just plumbing and positioning.

The architecture

You can build v1 in a weekend. The stack that keeps coming up in this space:

# routing layer — the whole moat is quota arbitrage
PROVIDERS = [
    gemini_cli_free,      # 1,000 req/day, primary
    workers_ai,           # 10,000 Neurons/day, overflow
    paid_flash_key,       # cents, last resort
]

def answer(context: bytes, question: str) -> str:
    for provider in PROVIDERS:
        if provider.quota_left():
            try:
                return provider.complete(prompt(question), images=[context])
            except RateLimited:
                continue
    return local_ollama_fallback()   # qwen2.5:3b on the user's own machine

Around it:

  • Overlay shell: a Tauri or PyQt frameless window, always-on-top, click-through except when summoned. Hotkey via pynput or Tauri’s global shortcut plugin.
  • Capture: mss for region screenshots on a 500ms rolling buffer; a vision model call fires only when the hotkey is pressed, so you’re paying tokens for one frame, not a video stream.
  • Audio mode: faster-whisper transcribes meeting audio locally (free, runs on CPU), feeding a rolling summary into the prompt so the model knows what’s been said — this is the actual “AI notetaker” feature, and it doubles as your legal camouflage.
  • Quota router: track each provider’s remaining quota per account in Redis or SQLite. Rotate. This is boring plumbing, and it’s exactly the part that determines whether your margin is 95% or negative.

Now the uncomfortable part: getting paid

Here’s the problem the Medium posts skip. A product marketed as “cheat on your interviews” is:

  • A ToS violation on every interview platform (CoderPad, CodeSignal, HireVue all ban it explicitly), every university’s academic code, and most employer policies.
  • Something payment processors will drop you for. Stripe’s restricted businesses list covers tools “designed to enable cheating”; one wave of chargebacks or one press inquiry and your account is frozen with the balance inside it. Ask anyone who’s sold grey-market subscription tools how that movie ends.

Cluely’s own history is the case study. The “cheat on everything” launch was a marketing grenade — 70,000 signups in a week, because outrage is free distribution. Then came the platform bans, the detection vendors, and the quiet pivot: today Cluely markets itself as an “AI meeting assistant” and prices a free plan with Pro from $11.99/month. The grey use case still drives word-of-mouth; the compliant packaging is what keeps the money flowing and let a16z write a $15 million check without blinking.

If you build this, learn the pattern instead of the aesthetics:

  1. Ship the compliant core: meeting notes, sales-call summaries, live research for support agents. Nobody’s ToS prohibits a notetaker.
  2. Let positioning do the grey work. Landing page copy, TikTok demos, the name — that’s where the “undetectable” energy lives. The code stays neutral.
  3. Collect money the way grey products collect money: crypto (USDT on a cheap chain) for the resale/proxy market, plus a mainstream processor for the compliant tier. Never let your only rail be one that can freeze on a Friday.
  4. Price per seat, monthly, $15–30. This is a subscription product, not a one-off gig. Fifty users at $20 is $1,000/month recurring on infrastructure that costs less than a coffee — and unlike “sell a service to local businesses,” fifty users is a Telegram group and a demo video away, not fifty sales calls.

The realistic ceiling for a solo builder is not Cluely’s $20M. It’s a few hundred users in a language and niche the big players ignore — interviews in Bahasa Indonesia, medical licensing exams, foreign-language sales calls — at $2,000–8,000/month recurring. That’s a real income from a product whose marginal cost is roughly zero.

The risks, stated plainly

  • Your users bear the ban risk. Getting caught using an overlay in a monitored interview ends careers, not accounts. A defensible product ships with honest warnings and no promises of invisibility — the arms race (that 38.5% detection rate) means today’s invisible window is next quarter’s flagged session.
  • Provider risk is real. Free tiers are marketing budgets. Google has cut Gemini CLI free limits before (Pro models went paid in March 2026); quotas in this article are October 2026 numbers and will move. The quota router with a local-model fallback is your insurance.
  • The ARR number that launched a thousand clones was inflated. Roy Lee admitted publicly in March 2026 that the famous “$7M ARR” was a lie — screenshots showed $5.2M. The market is real; the hype multipliers around it are not. Discount every revenue screenshot you see by half, including ones in your own head.
  • This is educational, not a business guarantee. Platforms patch, processors deplatform, detection improves. Build the compliant core, treat the grey demand as distribution, and keep your infrastructure costs near zero so a ban costs you a weekend, not your savings.

The Cluely story proves something worth more than the product itself: the gap between “AI that watches my screen and whispers answers” and “AI notetaker” is pure positioning — and positioning is free. The compute is already free. What’s left to build is the window.

Verified against: Anthropic pricing docs (web search $10/1,000 searches), Cloudflare Workers AI pricing docs (10,000 free Neurons/day, $0.011/1,000 overage), Gemini CLI free-tier limits (1,000 requests/day via Google account, 250/day via unpaid API key), Cluely’s funding and revenue history via Wikipedia and TechCrunch, and HeroHunt.ai’s interview-cheating analysis (38.5% of 19,368 interviews flagged). All limits current as of October 2026 — re-verify before you build, because none of these numbers signed a contract with you.