What it is
Cluely works by placing a transparent AI-powered layer above the candidate's active window, marking itself invisible to screen-capture APIs, then feeding real-time responses from a cloud LLM it queries over the internet. It popularized undetectable AI assistance during live technical and behavioral interviews, launching at roughly $20/month and spawning a wave of paid and open-source clones within months. On Windows, the specific mechanism is an OS-level flag called WDA_EXCLUDEFROMCAPTURE: any window with that flag set is omitted from the frame buffer a screen-share or recording actually reads, even though it still renders normally on the candidate's physical display. Cluely sets that flag on its overlay window, which is the entire trick.
Why it matters
Cluely turned undetectable AI coaching from a niche hack into a mainstream consumer category, alongside paid competitors like Interview Coder, Parakeet AI, and Final Round AI, and free open-source forks (OpenCluely, Pluely, Natively, MindWhisperAI) that anyone can compile under a different name in under an hour. Conventional proctoring tools operate at the browser or camera layer; Cluely and its imitators operate at the OS layer. These surfaces don't intersect, and no amount of watching harder closes the gap: a lockdown browser only controls the single browser window it launches, while Cluely runs as an entirely separate OS-layer process outside that sandbox, so it never enters the lockdown tool's field of view to begin with. Webcam proctoring misses it from the opposite direction, watching the candidate's face and room rather than screen contents, so an overlay that changes nothing about visible behavior gives the camera nothing to flag. The category demonstrated that legacy proctoring stacks cannot see overlay-class tools, and that no single named product needs to survive for the underlying technique to keep spreading. Aiseptor's dedicated Cluely detection page catalogues these signatures in more depth.
How Aiseptor addresses it
Aiseptor treats Cluely and every fork of it as part of a single class of threat: processes on the candidate device that must never reach their backends or render over exam content. Rather than matching a process name, Aiseptor targets the invariant every overlay of this kind is forced to expose: the WDA_EXCLUDEFROMCAPTURE-style exclusion flag itself, which cannot be removed without losing the invisibility that makes the tool worth using. Alongside that flag, Aiseptor enumerates top-most and overlay windows and running processes and watches for GPU VRAM deltas consistent with a loaded local model, both of which are observable at the OS level even though the overlay window is excluded from any capture API. Cluely specifically depends on a live connection to the OpenAI API to generate answers; when that endpoint, and the equivalent endpoints its forks call, is unreachable under a default-deny network policy, the overlay renders no answers regardless of how the binary is named or how it marks itself invisible. The enforcement runs on the network path via DNS/SNI-level blocking plus the OS-level technique signals, not on any specific binary name or process signature, which is why it catches renamed builds and forks it has never seen the same way it catches the original.
Definition
Canonical definition
Cluely is a commercial desktop application, launched in 2024 by founder Roy Lee, that displays AI-generated answers on a transparent overlay positioned above the candidate's active interview window. It pairs a local speech-to-text capture of the interviewer's questions with a cloud language model (Cluely calls the OpenAI API) to produce answers in real time, priced at roughly $20/month. By marking its surface as excluded from screen-capture APIs, the overlay is invisible to Zoom, Teams, Google Meet, and conventional screen recorders. Cluely's public traction, including venture funding, social-media-viral interview clips, and a wave of imitators, made it the reference tool for the invisible-overlay category. It is no longer alone there: paid competitors include Interview Coder ($100 lifetime), Parakeet AI ($20-40/month), Ultracode AI ($899 lifetime), LockedIn AI ($55-70/month), and Final Round AI ($149/month), while free open-source forks (OpenCluely, Pluely, Natively, MindWhisperAI) inherit its architecture and can be recompiled under an arbitrary process name in under an hour by anyone with basic developer skills. A motivated candidate can even build a comparable tool from scratch: Tesseract OCR, Whisper speech-to-text, and a Gemini API call reproduce Cluely's core capability in about two hours, with no signature for any detector to match. This reproducibility is what makes process-name-based detection architecturally unreliable at any point in time, and it explains the second wave of tools the category produced: process-name-scanning detectors (Honrly, Truely, Proctaroo) emerged in 2025 as a direct response to Cluely's viral moment, flagging known binaries and signatures. Each is defeated the same way Cluely's original detectors were: compile any open-source fork under a new name and the scan finds nothing, typically in under an hour. Cluely's own public positioning has leaned into rather than away from the controversy: founder Roy Lee gave a TechCrunch interview in July 2025 that embraced the controversy rather than distancing the company from it, treating detection tools built to catch it as a marketing foil rather than an existential threat, a stance that only holds up because its architecture, and that of its entire imitator category, genuinely has no reliable countermeasure at the application layer. Cluely is the reference example of what this site's glossary calls the invisible AI overlay category, an architecture pattern rather than a single product, and its founder has also floated a hardware roadmap, including earpieces and smart-glasses integration, that would move the same invisible-assistance model off the screen and into the physical interview room.
In practice
A concrete example
A detection vendor releases a tool that scans running processes for known Cluely signatures and flags matches to the proctor. Within days of the release, forum threads document how to recompile one of Cluely's open-source forks under a different process name and binary signature, defeating the new detector entirely with no change to the tool's actual capability. The underlying overlay technology never had to change; only the label on the process needed to. This is the recurring pattern that makes Cluely and its forks a category, not a single product a single detector could ever permanently retire. In a behavioral interview, a candidate is asked to describe a time they resolved a conflict with a coworker. Cluely's audio pipeline captures the question as it's spoken, sends it to the LLM, and renders a structured STAR-format response (Situation, Task, Action, Result) on the transparent overlay within about a second, specific enough to include invented but plausible metrics. The candidate reads it verbatim while appearing to recall a real memory, and nothing about the screen, the webcam feed, or the audio the interviewer hears reveals that the answer was generated rather than recalled.
Citations
- [1]Fabric, analysis of 19,368 AI-conducted interviews, January 2026 (2026)
- [2]Talview AI Threat Index Report 2026 (2026)
- [3]TechCrunch interview with Cluely founder Roy Lee (2025)