Cluely & Cheating Detection 2026: Real-Time AI Overlays Tested

Cluely is a real-time AI overlay assistant that watches your screen, listens to your microphone, and whispers answers into a window your Zoom participants supposedly can’t see. What started as an interview-cheating tool built by a suspended Columbia student has become an entire category, complete with an open-source clone called Glass and a dozen “live sales copilot” startups selling the identical binary with a nicer landing page. In mid-2026 the platforms fought back: Zoom, Google Meet, and the major proctoring vendors all shipped screen-share integrity checks that specifically hunt for invisible overlay windows. If your sales team is about to deploy one of these, you need to understand what you’re actually installing.

What’s actually new with the real-time AI overlay assistant category

The technical trick behind every tool in this space is the same, and it’s the thing that’s now breaking. On macOS, a window can set NSWindowSharingNone, which tells the compositor to exclude it from screen capture. On Windows, the equivalent is SetWindowDisplayAffinity with WDA_EXCLUDEFROMCAPTURE. Both were designed for legitimate reasons — DRM playback, password managers, banking apps that keep credentials out of screenshots. Overlay assistants repurposed those flags so the assistant panel floats above your video call, visible to you, absent from the shared stream. Pair that with continuous screen OCR and system-audio transcription piped into a fast model, and you get answers roughly two seconds after a question is asked out loud.

What changed in 2026 is that the capture-exclusion flag is no longer a reliable invisibility cloak. Conferencing platforms now cross-reference what they capture against what the OS reports is on screen. If the window manager lists eleven visible windows in the shared display region and the captured frame accounts for ten, that delta is a signal. Zoom’s integrity checks flag it to the host; several proctoring vendors treat it as a hard fail. Google Meet added process-level heuristics for known overlay binaries. None of this is airtight — a determined user can run the assistant on a second device with a camera pointed at the screen, which no software check can see — but the era of “install it and you’re invisible” is over for anyone using stock builds.

The more interesting development is commercial. Cluely pivoted hard from “cheat on everything” to enterprise sales enablement, and the copycats followed. Glass, the open-source Cluely alternative, sits on GitHub under a permissive license, so any vendor can fork it, restyle it, and sell it as an invisible AI meeting assistant for sales teams. When a rep pitches you an AI screen and audio copilot, there is a real chance the underlying code is the same project that got someone fired from a job interview last year — with the same capture-evasion flags still compiled in.

Why it matters

• Your compliance posture may already be broken. A tool that transcribes system audio is recording every participant on the call. In two-party consent states and under GDPR, silent recording without disclosure is a legal exposure, not a feature. The invisibility that makes the tool appealing is precisely what makes the consent problem unfixable.
• “Undetectable” is now a liability claim, not a selling point. If a vendor’s marketing leans on evading Zoom or Google Meet, you’re buying a product on a collision course with the platforms you depend on. Overlay detection is a moving target, and the vendor is one update away from a dead feature.
• Screen OCR means total data capture. These tools read whatever is on screen — open CRM records, a Slack DM, a customer’s payment details in a support ticket — and ship it to a model endpoint. Ask where that data lands and how long it’s retained before anyone installs it.
• Hiring pipelines are measurably degraded. Take-home tests and unproctored video interviews are now low-signal. Companies that haven’t restructured technical screening are selecting for tool access, not skill.
• The legitimate use case is real and narrow. A live sales call AI copilot that surfaces pricing, objection handling, and competitor intel to your rep in real time is genuinely valuable. The value comes from retrieval over your own content, not from hiding the window.
• Shadow IT risk is high. These are consumer-grade installs with a credit card and a download. Individual reps will deploy them without procurement, and the first you’ll hear about it is from a customer who noticed.

How to use a real-time AI overlay assistant today (without getting burned)

1. Audit what’s already installed. Before evaluating anything, find out what’s running on company machines. On Windows, check for the common overlay binaries and any process requesting capture exclusion:

   Get-Process | Where-Object { $_.ProcessName -match 'cluely|glass|pickle|interview|cheat' } |
     Select-Object ProcessName, Id, Path, StartTime

   On macOS, list running apps and look for anything holding an accessibility or screen-recording permission you didn’t approve:

   sudo sqlite3 /Library/Application\ Support/com.apple.TCC/TCC.db \
     "SELECT client, auth_value FROM access WHERE service='kTCCServiceScreenCapture';"

2. Run the open-source version first, in a sandbox. Glass gives you the whole category’s behavior for free, which makes it the honest way to evaluate before you sign a contract. Run it on a test machine with no production credentials:

   git clone https://github.com/pickle-com/glass.git
   cd glass
   npm install
   npm run setup   # prompts for your model API key
   npm start

   Watch what it sends. Point a proxy at it and read the payloads — you will learn more in twenty minutes than from any vendor security questionnaire.

3. Set the disclosure policy before the pilot, not after. If a copilot is on the call, participants get told. Put one line in the meeting invite template and make it non-negotiable:

   This call may use an AI assistant that transcribes audio to
   generate notes and surface reference material for our team.
   No recording is retained after 30 days. Reply to opt out and
   we'll run the call without it.

4. Point it at your own content, because that’s where the value is. A generic model guessing at your pricing is worse than useless. Whatever tool you pick, load your objection-handling docs, pricing sheet, and competitor battlecards into its context. A minimal system prompt for a sales copilot:

   You are a live sales copilot listening to a customer call.
   Output at most 2 short bullets at a time. Never output filler.
   
   Trigger on: pricing questions, competitor mentions, technical
   objections, buying signals ("when could we start", "who else
   uses this").
   
   Answer ONLY from the attached battlecards and pricing sheet.
   If the answer isn't there, output: "NOT IN DOCS — don't guess."
   Never output anything the rep should not say out loud verbatim.

5. Disable capture evasion where the tool allows it. Several vendors now expose a visible mode. If yours doesn’t, treat that as disqualifying. In Glass-derived forks the flag is set at window creation, so a self-hosted fork can strip it:

   // electron main process — make the overlay honest
   const win = new BrowserWindow({ transparent: true, frame: false })
   // REMOVE this line in a compliant build:
   // win.setContentProtection(true)
   win.setContentProtection(false)

6. Test against your actual stack. Run a real Zoom and Google Meet call with the tool active and a colleague hosting. Ask whether the host dashboard flags anything. Do this monthly — platform integrity checks change without notice, and a feature that worked in the pilot can trip an alert in production.

7. Fix the interview process separately. If you hire technical staff, stop treating unproctored screening as signal. Move to live pair-programming on a shared repo, ask candidates to explain code they didn’t write, and interrupt with follow-ups. Overlay tools answer questions well and defend answers badly.

How it compares

Tool	Model	Positioning	Capture evasion	Best for
Cluely	Paid subscription, hosted	Sales/support copilot (pivoted from interview aid)	Yes, core feature	Teams wanting a polished product and willing to accept vendor data handling
Glass	Open source, self-hosted	General-purpose overlay assistant	Yes, toggleable in a fork	Evaluating the category, or building a compliant internal build
Otter / Fireflies	Paid SaaS	Visible meeting notetaker	No — announces itself	Compliance-safe transcription and post-call summaries
Gong / Chorus	Enterprise contract	Revenue intelligence, post-call coaching	No	Sales orgs that need analytics and manager review, not live prompts
Native platform AI (Zoom AI Companion, Gemini in Meet)	Bundled with seat	In-platform summaries and live assist	No — first-party	Lowest-risk starting point; already covered by existing DPAs

The honest read: if you want the business outcome — reps who don’t fumble pricing questions — the native platform assistants and a well-maintained battlecard get you most of the way there with none of the legal exposure. The overlay category earns its keep only when you need sub-two-second retrieval over proprietary content that no first-party tool can index.

What’s next

Expect the OS vendors to close the loop the conferencing platforms opened. Capture exclusion was never meant to be adversarial, and both Apple and Microsoft have obvious levers: restrict the flag to notarized applications with a declared entitlement, or surface a persistent system indicator whenever a window is excluded from an active capture session. The moment an OS-level badge appears next to the screen-share indicator, the entire “invisible” value proposition evaporates and the category has to compete on retrieval quality alone. That would be healthy.

On the enterprise side, watch procurement language catch up. Security questionnaires now include direct questions about screen OCR scope, audio capture consent, and content-protection flags. If you buy software, add those three questions to your own template — they separate a legitimate real-time AI overlay assistant from a repackaged cheating tool faster than anything else. Vendors who answer cleanly are building for the enterprise; vendors who dodge are selling invisibility.

The third thing to watch is the local-model shift. Overlay tools are latency-bound, and the sub-second small models now running on consumer laptops handle retrieval-and-summarize work well enough. When the whole loop runs on-device, the data-exfiltration objection largely disappears — genuinely good news for the compliant use case, and also the removal of the last technical constraint on the non-compliant one. The tooling isn’t going away. The only question is whether your organization deploys it deliberately or discovers it accidentally.

Frequently Asked Questions

Can Zoom or Google Meet actually detect an invisible AI meeting assistant?

Increasingly, yes — but not universally. The 2026 integrity checks compare the captured frame against what the window manager reports is on screen and flag the discrepancy to the host. Known binaries get matched by process heuristics. What no software check can catch is an assistant running on a separate phone or tablet, which is why proctoring vendors have moved toward camera-based room monitoring.

Is it legal to use a live sales call AI copilot?

Using one is generally legal; recording the call without disclosure often isn’t. Any tool transcribing system audio is recording every participant. In two-party consent jurisdictions and under GDPR, you need disclosure and a lawful basis. Disclose in the invite, honor opt-outs, and set a retention limit. That single practice converts most of the legal risk into a policy question.

What’s the difference between Cluely and Glass?

Functionally, less than the pricing suggests. Glass is the open-source Cluely alternative — same overlay approach, same capture-exclusion flags, self-hosted with your own API key. Cluely is a hosted commercial product with a polished interface, managed updates, and a vendor to hold accountable. Glass is the better evaluation tool because you can read the code and watch the network traffic.

Should I ban these tools at my company?

Ban unsanctioned installs, not the capability. A blanket ban produces shadow IT — reps will install it personally and you’ll lose visibility entirely. Sanction one tool, require visible mode, mandate disclosure, and log what it captures. That gets you the productivity benefit with an audit trail.

Do these tools work well enough to matter?

For retrieval, yes — surfacing a pricing tier or a competitor comparison two seconds after it’s mentioned genuinely helps a rep. For reasoning, they’re mediocre. They produce confident answers to questions requiring context they don’t have, and a rep reading one aloud will get caught by any customer who asks a follow-up. Constrain them to retrieval from your own documents and they earn their cost.

How do I stop candidates from using an AI screen and audio copilot in interviews?

You mostly can’t, so change what you’re measuring. Live pair-programming, asking candidates to critique code rather than write it, and aggressive follow-up questioning all defeat overlay tools, because the tools answer the asked question and can’t defend the answer. Proctoring software is an arms race you’ll lose; interview design is a fix you control.