Cluely Teardown — The 2025 Viral "Cheat on Everything" AI ($1M+ MRR, College Dropouts, a16z-Backed)

TL;DR

Cluely is a desktop overlay that reads your screen, listens to your microphone, and whispers AI-generated answers into your ear during interviews, sales calls, customer support, dates, and meetings. Roy Lee, 21, built the early version to cheat on his Amazon technical interview. Amazon caught him. Columbia expelled him. He posted the cease-and-desist on Twitter and turned the punishment into a product launch. A few months later, the company had a16z money, a co-founder (Neel Shanmugam, also expelled), a viral launch video that drew over 6 million views in week one, and a run-rate that public estimates pegged somewhere between $1M and $3M MRR by mid-2025.

The product itself is not technically remarkable. Whisper for transcription, GPT-4o and Claude for completion, Electron for the overlay, screen-share APIs for capture. A solo developer with decent taste could build a working prototype in a long weekend. What is remarkable is the distribution. Cluely is a teardown of one specific thing: how two 21-year-olds turned getting expelled into a $5.3M Series A and a thousand imitators.

You cannot copy the product. The product is a commodity. You can copy the playbook, and the playbook is in section 7.

Quick Facts

Field	Value
Founded	Early 2025 (initial cheating tool ~Feb 2025, rebrand to Cluely ~Apr 2025)
Founders	Roy Lee, Neel Shanmugam (both Columbia CS, both expelled spring 2025)
Funding	$5.3M seed led by a16z + Susa Ventures + Pioneer Fund, mid-2025
Revenue	Public estimates $1M-$3M MRR by mid-2025; founder claims $4M ARR within months
Pricing	$20/mo Pro (individual), Team and Enterprise tiers added later
Platform	macOS, Windows (Electron overlay)
Audience	Sales reps, customer support, job seekers, students, knowledge workers
Position	Anti-establishment, anti-AI-safety, "cheat on everything"

These numbers are a moving target. Founders inflate, journalists round, and "ARR" in 2025 sometimes means "the number we got one good month and multiplied by 12." Treat the precise figure as less important than the shape: small team, viral launch, real paying users, real revenue at a scale that would have taken a 2018 SaaS startup three years to reach.

A 5-Minute Walkthrough

You install the app. It asks for screen recording permission and microphone permission. Both are required. The app sits invisibly on top of every window. There is no taskbar icon, no menu bar entry. Press a hotkey and a small chat-window slides in from the side of your screen. It already knows what is on the screen and what was said in the last few seconds, because it has been watching and listening the entire time.

You join a Zoom call. The other person asks you a technical question. Cluely transcribes them in real time, runs the question through GPT-4o, and types the answer into the overlay before they finish speaking. You read the answer off your screen and say it. The overlay is invisible to the other party because it lives at the OS compositing layer, above the screen-share frame.

You open LeetCode during an interview. The overlay reads the problem, generates a solution with comments, and walks you through the explanation. The interviewer cannot see the second window because there is no second window.

You are on a sales call. The prospect says they are worried about pricing. The overlay surfaces the three best objection-handling scripts your team has saved, ranked by the prospect's actual phrasing. You pick one and say it.

You are on a Hinge date that is going badly. The overlay listens through your phone microphone and suggests a topic change.

The product is, in literal terms, an always-on second brain that runs alongside any conversation you are having. The marketing is not subtle about this. The tagline is "Cheat on everything." The launch video showed Roy on a date using Cluely to fake interest in art history. The video closes with him getting caught and then shrugging.

What you notice when you actually use it: the latency matters more than the model. Whisper streaming has a roughly 800ms-1.5s delay from speech to text. GPT-4o adds another second or two. That means the overlay's suggestion lands roughly two to three seconds after the other party finishes their sentence — which is the awkward beat where a real human would already be answering. You learn to use it for the second half of an answer, not the first. Sales reps in the Discord report this same thing. Customer support agents handling chat (not voice) report it works better because text gives you more time.

Business Model

Cluely runs a simple consumer-SaaS price ladder: $20/mo Pro for individuals, custom Team pricing, custom Enterprise pricing. The individual plan is the loss leader and the marketing engine. The team plan, which started as an afterthought, is reportedly where the revenue concentration is shifting in the second half of 2025 — sales orgs paying for ten or fifty seats at a higher per-seat number.

The unit economics on the individual plan are tight. Each active user runs Whisper + GPT-4o + Claude inference for several hours a day. Even at OpenAI's bulk pricing, a heavy user can burn through $3-$7 in API costs per month against $20 in revenue. The gross margin on a casual user is healthy (60-70%) but a power-user sales rep can take the margin under 30%. Cluely has been rolling out usage caps and tier upgrades to fence this in.

The competitive moat is not technical. There is no proprietary model, no data flywheel, no network effect. The moat is brand. Cluely owns the position of "the cheating AI." A team that wants to ship the same product in 2026 has to either pick a different position (boring) or fight Cluely on its own ground (expensive). Roy Lee's Twitter following — over 200k by mid-2025 — is the actual asset. The Electron binary is just the thing the brand sells.

The retention number nobody talks about: this is a tool people churn off the moment the immediate use case ends. Get the job, churn. Close the deal, churn. Finish the semester, churn. Cluely's growth right now papers over the churn because the top of funnel is wide. The interesting question for 2026 is whether they can build a use case that creates daily habit (meetings, support, sales) before the viral moment cools.

Tech Stack

The stack is short. This is part of the lesson.

• Frontend: Electron app (TypeScript + React) for the overlay. Native macOS and Windows builds. Uses OS-level screen capture APIs and microphone access. The overlay is rendered at the compositor level so it is invisible to screen sharing — that piece is the one bit of nontrivial engineering in the entire product.
• Speech-to-text: Whisper, streamed. Some reports suggest they have moved part of the pipeline to Deepgram or a fine-tuned whisper.cpp for latency.
• LLM: GPT-4o for most completions, Claude (probably Sonnet) for longer-context tasks, with model routing based on use case.
• Backend: Standard Postgres + Redis + a Node or Python API. Stripe for payments. Auth via standard email/password and Google OAuth.
• Hosting: AWS, with some inference traffic going direct to OpenAI/Anthropic.

A capable solo developer could rebuild a feature-equivalent version in two to four weeks. The screen-capture-invisible-to-screen-share trick is the only piece that requires non-obvious knowledge, and there are open-source examples of it on GitHub. The transcription, the LLM call, the streaming UI, the Stripe wiring — all of it is undergraduate territory in 2025.

This is the point. Cluely is not a moat made of code. The code is a commodity that the brand uses as a delivery mechanism. Every dollar of Cluely's valuation is paying for Roy Lee's distribution, not for engineering.

The Controversy Playbook

This is the section that matters. If you are reading this teardown to learn how Cluely actually got built, skip everything else and read this carefully. The technical product is replaceable in a month. The distribution playbook is the thing other founders are studying.

Step 1: Pick a real workflow people already do quietly, then say the quiet part loud.

Job candidates have used unauthorized help for as long as there have been job interviews. Salespeople have always had cheat sheets. Students have always cribbed. The interesting move is not to invent a new behavior. It is to find a behavior most people already do, and remove the shame. Roy Lee did not invent interview cheating. He took an existing widespread behavior, gave it a $20/mo product, and made it a personality.

The replicable insight: scan your social circle for things smart people do but feel bad about. Hinge openers written by ChatGPT. Performance review self-evaluations written by Claude. Cold email replies written by Lex. None of these are secrets. All of them are embarrassments. Embarrassments convert to products if you have the stomach to put your name on the position.

Step 2: Use the punishment as the launch.

Roy Lee was caught cheating in an Amazon interview. He could have apologized, gone quiet, transferred schools, and shipped the same product six months later under a neutral name. Instead he posted the cease-and-desist letter, the Columbia expulsion paperwork, the recording of the Amazon interviewer realizing what was happening, all on Twitter, with no apology and a clear "they are mad because they are losing" frame.

The expulsion was the marketing event. Without the expulsion, Cluely is a tool. With the expulsion, Cluely is a story. Stories scale. Tools do not.

Step 3: Pick a tagline the establishment cannot say back.

"Cheat on everything" is not a tagline a venture-backed startup is supposed to use. It is the kind of phrase that gets workshopped out by a brand consultant in week one. It is also the entire reason Cluely cuts through.

The Coca-Cola test: would the CEO of a Fortune 500 company put your tagline on a slide in front of the board? If yes, your tagline is too safe.

Step 4: Pre-build an enemy.

Cluely has explicit, named enemies. AI safety researchers ("doomers"). Corporate AI guidelines. The big-tech recruiting process. Universities.

Enemies do two things. They give your supporters something to be against, which is more energetic than something to be for. They also produce an opposing-counter-attack on Twitter, which is more press.

Step 5: Hire a face.

Roy Lee is the company. Not the CEO, not the founder, the company. He is on every podcast, every video, every Twitter thread, every TikTok. His face is the brand.

Step 6: Stage your moments.

The "Hinge date" launch video was not a candid recording. It was scripted, lit, edited, and timed for a Tuesday morning Twitter drop. Cluely's content output looks unhinged but is operationally tight.

Step 7: Get the funding to validate the brand, not the product.

The a16z check did one specific thing: it took Cluely from "scam app made by an expelled kid" to "venture-backed company that an expelled kid is running." Same product, same founder, same controversy, different frame.

Step 8: Plan for the rotation.

The "cheat on everything" brand has a half-life. Cluely has roughly 12-24 months before the controversy becomes background noise. The team has signaled the second act is enterprise — sales orgs paying $200/seat/mo for compliant versions of the same product.

Why Now

Three things had to be true for Cluely to exist in 2025 and not in 2022. First, LLMs had to be good enough to give useful real-time answers. Second, the cultural Overton window on AI had to shift from "doomer" to "accelerationist." Third, the cost of inference had to drop low enough.

If you are reading this in 2026 and wondering whether the window is still open: yes, but not for "Cluely for X" copycats. The window is open for new behaviors that were shameful before AI made them frictionless.

Founders

Roy Lee is the public face. 21, Columbia CS dropout (involuntary), Korean-American. Before Cluely he had a YouTube channel posting Leetcode walkthroughs — that's where his initial audience came from.

Neel Shanmugam is the co-founder. Less public, more operational. Also Columbia CS, also expelled. Roy is product, marketing, fundraising, face. Neel runs engineering, hiring.

A few honest things: they are young, have not run a company before, have made one product that worked. The hard part — turning a viral consumer brand into durable enterprise revenue while maintaining the original brand energy — is in front of them.

The thing they have already proven is that two college dropouts with no industry connections can ship a brand-driven consumer-SaaS product to $1M-$3M MRR in a few months in 2025.