Cluely is an AI-powered meeting assistant designed for complete undetectability, offering real-time assistance, automated note-taking, and pre-meeting insights. Its core methodology emphasizes local processing to remain invisible to other meeting participants, unlike conventional AI notetakers that join as visible bots.

The system's real-time question-answering capability functions by integrating three primary data streams: the user's screen content, the live audio transcript, and an underlying artificial intelligence model, presumably a large language model (LLM) or a similar generative AI. Technically, this involves continuous optical character recognition (OCR) on the screen display to capture visual context ($C_{screen}$), alongside real-time speech-to-text (STT) processing of the audio feed to generate a conversational transcript ($T_{audio}$). These contextual inputs are then fed into the AI engine, which responds to user queries, e.g., "What should I say?", by synthesizing information from the combined context: $Response = AI_{model}(Query, C_{screen}, T_{audio})$. This real-time processing boasts a transcription response time of 300 ms and an accuracy of 95%, supporting over 12 languages.

For meeting notes and follow-up emails, Cluely's AI engine analyzes the comprehensive record of the meeting, which includes both the detailed audio transcript and the visual information gleaned from the screen. The notetaking process involves a multi-modal summarization task, where the AI distills key points, decisions, and action items from $T_{audio}$ and $C_{screen}$ to produce structured, shareable notes. Similarly, the generation of follow-up emails leverages this consolidated meeting data to draft relevant communications instantly.

Participant insights are generated pre-call by accessing and synthesizing information from external data sources (e.g., professional networks, company databases) based on known participant identities, providing the user with context on "who they are really talking to" before the meeting commences. The system is initiated and stopped directly by the user, implying a client-side application that captures audio and screen data without requiring integration as a virtual participant in the conferencing software itself, ensuring its "undetectable" claim by operating solely from the user's local environment.