How to detect Cluely in interviews: what works and what doesn't

If you interview remotely, you have probably wondered whether a candidate was using Cluely. The honest answer is that you are unlikely to see Cluely itself. What you can see is what it does to an interview, and you can run interviews in a way that makes hidden help far less useful. This guide covers both, and it is just as relevant to the other AI interview copilots that work the same way.

What Cluely is and why video calls can’t see it

Cluely is a desktop app that listens to a call, reads what is on the screen and uses large language models to suggest answers in real time. Wikipedia describes its overlay as “designed to be invisible during screen-sharing”. Two of its founders had earlier built Interview Coder, a tool for solving coding interview questions. By November 2025, Cluely had repositioned itself as a general AI meeting assistant and removed explicit references to cheating on exams and interviews from its website.

The rebrand does not change the problem for hiring teams. Any tool that listens to the question, writes an answer and keeps its window out of the screen share raises the same issue.

Why can’t you see it? Operating systems let an app mark its own window so that it is left out of screen capture. Microsoft documents one such setting for Windows: with it, “the window is displayed only on a monitor. Everywhere else, the window does not appear at all.” Microsoft’s own example is hiding recording controls from a recording. An app that uses this kind of setting shows up on the candidate’s monitor but not in what Zoom, Google Meet or Teams send to you. You see their editor. They see their editor plus the suggested answer.

Can Cluely be detected? An honest answer

It helps to separate two things: detecting the app, and detecting the behaviour.

Detecting the app by name needs software that runs on the candidate’s computer and can inspect its processes and windows. The vendors we found that claim to catch named tools use installed software. InterviewGuard describes an installed candidate app that monitors at system level. Zero Assist describes an operating-system agent that checks running processes against a list of known cheating tools. Codility offers a desktop app that detects unauthorised applications, on its Custom plan only (all as of October 2026). Even installed agents are in an arms race, because the tool makers update their products to stay hidden.

Detecting behaviour is what a video call or a web page can do. It cannot list the other apps on a computer. It can see what happens inside the session: whether the candidate stays in fullscreen, switches tabs, moves the cursor away, or copies and pastes text.

Eagle Vision is in the second group. Candidates install nothing: they open a link in a desktop browser. It does not identify Cluely or any other app by name. It makes the behaviours around hidden help visible and recorded, and leaves the judgement to the interviewer.

Be realistic about the limits, too. Some copilots now market ways around behaviour checks. Parakeet AI’s website lists “invisible to tab switching” and “cursor undetectability” among its features. No single control settles the question. Interview technique does the rest of the work.

The behaviour overlay copilots leave behind

A copilot has to hear the question, generate an answer and show it to the candidate, who then reads it out. That pipeline leaves traces.

• Echoing the question. The candidate repeats or paraphrases your question before answering. It buys time while the tool transcribes and writes.
• The same delay before every answer. Fabric, which sells AI interviews, describes a steady three-to-five-second pause regardless of how hard the question is. People normally answer easy questions faster than hard ones.
• Eyes fixed on one spot. Reading looks different from thinking. Watch for steady side-to-side scanning near the camera, rather than the glances up and away that come with recall.
• Polished structure, no false starts. Answers that open with “There are three main considerations” and run on in perfect paragraphs, with none of the restarts real speech has.
• Vocabulary they can’t unpack. A precise technical term, followed by a vague answer when you ask what it means.
• In coding rounds: code that arrives in complete blocks, no visible exploration or debugging, and trouble explaining a line they have just written.

Any one of these has innocent explanations. Nervous candidates echo questions. Some people keep the call window on a second monitor. Neurodivergent candidates may avoid eye contact, and a poor connection can create steady delays. Look for several signals together, repeated across the interview. Our guide to the signs a candidate is using AI covers this in more depth.

Interview Coder, Parakeet, LockedIn and Final Round: what’s different

The descriptions below come from each vendor’s own website, checked in October 2026. These products change quickly.

Interview Coder

A desktop app aimed at coding interviews. It markets itself as “100% invisible to screen-recording” and says it hides from the dock. Because it targets coding rounds, the best counter is to make candidates explain, modify and test their code live, rather than judging the final answer.

Parakeet AI

A real-time call assistant that listens, transcribes and generates answers. It is offered as a desktop app and in desktop and mobile browsers. If it runs on a phone propped beside the laptop, nothing on the computer will see it. The camera and your follow-up questions matter most here.

LockedIn AI

An assistant that listens to the interview and gives real-time answers, code and coaching, with a desktop “Background Mode”. Its “LockedIn Duo” feature brings in a human helper, and the vendor says “the interviewer sees only you and your normal screen share”. That puts it close to a proxy interview, so our guide to proxy interviews is relevant.

Final Round AI

Its Interview CoPilot streams short answers “in a private window only you can see”. The vendor says its Stealth Mode is on by default and that the windows are hidden from screen sharing and recording in Zoom, Google Meet and Teams. It also lists coding platforms such as HackerRank and CoderPad.

Interview design that overlays struggle with

Copilots are good at the first answer to a standard question. They are much weaker when the conversation gets specific, personal or fast.

• Change the constraints mid-problem. Once a solution works, add a memory limit, a new input format or a concurrency requirement, and ask for the change straight away.
• Make them explain their own code. Point at a line and ask why it is there. Ask them to predict the output for a new input before they run it.
• Ask about past decisions. “On your last project, what did you choose instead of X, and who disagreed?” Then ask about the people, the dates and what went wrong.
• Follow up on their own words. “You mentioned a cache. What was the eviction policy, and why?” A copilot answering the previous question is a step behind.
• Ask short questions quickly. Rapid, small follow-ups leave no time for a generate-and-read loop.
• Watch the process on a shared code pad. How someone breaks a problem down is harder to fake than the finished code.

Our guide on how to run a remote technical interview has a full structure you can reuse.

Lock the session so stepping out is visible

Interview technique works best when the session itself is controlled. Eagle Vision runs live interviews in fullscreen lockdown. During the interview, you see a flag as it happens when the candidate leaves fullscreen, switches tabs, moves the cursor to another screen, or copies, cuts or pastes text (copying inside the shared code pad or chat is not flagged). An integrity score out of 100 drops by a fixed amount with each flag, and you can watch it update during the call. You talk face to face on camera, write and run code together on a shared code pad, and get a recording and report afterwards.

To be clear about what this does: it flags behaviour, not app names. An overlay that never makes the candidate leave the interview screen may not trigger a flag at all. But many ways of using hidden help do involve stepping out: copying an answer from another window and pasting it, moving the mouse to a second monitor, or switching to a browser tab. Those become visible as they happen. Every flag is timestamped, so in the report you can click a flag and jump straight to that moment in the recording, then look for the behavioural signs above around it.

Tell candidates the rules up front

Before the interview, tell candidates what is monitored and recorded, and whether any AI tools are allowed. Some employers now allow AI in certain rounds, so do not assume candidates know your policy. Explain what happens if something is flagged: a human reviews the recording before any decision is made. Link to your privacy policy so candidates know how recordings are stored.

Clear rules deter most people, and they make any later conversation about a flag far easier. Our guide on how to prevent cheating in online interviews has more on setting up a fair process.