cheating-daddy: a real-time AI assistant for live calls, built on Electron and Gemini Live

What cheating-daddy actually does during a live call

The README describes cheating-daddy as a real-time AI assistant that provides contextual help during video calls, interviews, presentations and meetings. The mechanism it advertises is screen capture plus audio analysis: the app watches what is on your display and listens to what is being said, then produces an answer you can read from an always-on-top overlay window. The repository ships five profiles, listed in the README as Interview, Sales Call, Business Meeting, Presentation and Negotiation, which suggests the prompt and framing change with the scenario rather than the app behaving identically in all of them.

The intended user is someone who is already in the call and needs an answer faster than they can type a question into a chat window. The README is explicit that this is not a general note-taking tool: it says the assistant provides real-time assistance based on your screen and what the interviewer asks. The overlay, click-through mode and keyboard-driven window movement all exist so the window can sit on top of a video call without stealing focus or blocking clicks. That combination is the product. Everything else is plumbing.

How the Electron app, Gemini Live and the audio path fit together

The architecture visible in package.json is small. The single runtime dependency that matters is @google/genai, the Google Gen AI SDK, alongside ws for WebSocket work and electron-squirrel-startup for Windows installer handling. The main entry is src/index.js, and the app is packaged with Electron Forge, which is why the devDependencies are almost entirely makers: maker-dmg for macOS, maker-squirrel for Windows, maker-deb and maker-rpm for Linux, plus a community AppImage maker. The README names Google Gemini 2.0 Flash Live as the model, and the word Live matters: this is a streaming session, not a request-response call, which is what makes sub-second contextual replies conceivable at all.

Audio capture is the part with the most platform-specific handling, and the README is honest about the split. On macOS the app relies on SystemAudioDump for system audio. On Windows it uses loopback audio capture. On Linux it falls back to microphone input, which is a materially different feature: microphone input hears the room, not the call. That single line in the README explains most of why the Linux build is described as testing-only. If you are evaluating this project, the audio path is the thing to inspect in src/, because it determines whether the assistant hears the interviewer, the room, or nothing useful.

Installing cheating-daddy and running a first session

There is no published binary in the README, so installation means cloning the repository and building from source with Node and npm. The README gives three steps: get a Gemini API key from Google AI Studio, install dependencies, and start the app. The package.json scripts confirm the commands.

npm install
npm start

The start script maps to electron-forge start, so npm start launches the Electron app in development mode rather than installing a packaged build. The README also lists package, make and publish scripts, which are the Electron Forge commands for producing distributables, but no prebuilt release artifact is documented.

Once the window opens, the README's usage flow is: enter your Gemini API key in the main window, choose your profile and language in settings, then click Start Session. After that you position the window with the keyboard shortcuts. Window movement is Ctrl/Cmd plus the arrow keys, Ctrl/Cmd + M toggles click-through so the window stops intercepting mouse events, Ctrl/Cmd + \ closes the window or goes back, and Enter sends a message to the AI. The README adds a note that during testing the app will not answer if you ask it something; you need to simulate an interviewer asking a question, which it will then answer. Treat that as the expected interaction model rather than a bug.

Before any of this works you need screen recording and microphone permissions granted to the app, which the README lists under Requirements alongside an Electron-compatible OS and a Gemini API key.

Where cheating-daddy breaks or is the wrong tool

The Linux story is the clearest limitation, and it comes from the project itself. The README lists cross-platform support as macOS, Windows and Linux with the parenthetical "kinda, dont use, just for testing rn". Combined with microphone-only audio capture on Linux, that means the assistant cannot hear the far side of a call on that platform. If your work happens on Linux, this is not a rough edge you can configure around; the audio source is different.

The second limitation is the interaction model. The README states that the app will not answer when you ask it something and that you need to simulate the interviewer asking a question. An assistant that only responds to the other party is a poor fit for solo work, for reviewing your own recordings, or for any situation where you are the one driving the conversation. It is built for reacting, not for being asked.

The third is maturity. The latest release in the README and package.json is v0.8.0, dated 2026-07-23, following v0.7.0 and v0.6.0 in February 2026. A pre-1.0 version number, no test suite visible in the repository layout, and a lint script that is literally echo "No linting configured" all point to a project where you should read the source before trusting it with a live call. The README also does not document what happens to your captured screen and audio beyond being sent to Gemini, and it does not document rollback or data retention. If that silence matters to you, inspect src/ before you click Start Session.

cheating-daddy against Clue and other interview copilots

The obvious comparison is Clue, which is one of the alternatives people search for alongside this project. The structural difference is where the intelligence runs and who pays for it. cheating-daddy is a local Electron application that you run yourself, and it authenticates directly to Google with your own Gemini API key that you paste into the main window. That means the model cost lands on your Google account and the app itself is free under GPL-3.0. A hosted interview copilot typically bundles the model access into a subscription and keeps the capture pipeline on its own infrastructure, which removes the API key step but also removes your ability to inspect or modify the pipeline.

That trade cuts both ways. Self-hosting with your own key means no vendor between you and the model, and the GPL-3.0 licence means you can read and fork the code. It also means you own the setup: permissions, platform audio quirks, and the API key itself. The other difference worth noting is packaging. cheating-daddy ships through Electron Forge with makers for dmg, squirrel, deb, rpm and AppImage configured in package.json, but the README points users at npm install and npm start rather than a download page, so the practical path today is building from source. If you want a signed installer and a support channel, this is not that.

Licence, maintenance and the cost of upgrading

cheating-daddy is licensed GPL-3.0, which is a copyleft licence: if you distribute a modified version, the source of your version has to be available under the same terms. For an individual running the app locally this changes nothing. For a team that wants to fork it, rebrand it, or embed the capture pipeline in an internal tool that gets distributed, the obligation is real and worth checking with someone qualified. This is a description of the licence, not legal advice.

The repository is not archived, and the last push was on 2026-07-23, which is also the date of v0.8.0. The gap between v0.6.0 and v0.7.0, both dated 2026-02-09, and then the jump to v0.8.0 in July suggests an irregular release rhythm rather than a fixed cadence. Upgrade cost is dominated by the Electron and @google/genai versions in package.json: Electron is pinned at ^30.0.5 and the Gen AI SDK at ^1.2.0, so a major bump in either can require changes in the main process or in how the Live session is opened. Because there is no configured lint or test step, you cannot rely on CI to catch a regression after an upgrade; you would be verifying the capture and streaming path yourself.

Editorial conclusion

Use cheating-daddy if you are comfortable reading source and running an Electron app from a repository that is still pre-1.0, and you already have a Gemini API key. Do not use it on Linux, where the README says the build is for testing only, and do not expect it to answer when you speak first, since the README notes it responds to an interviewer's question rather than to you. Before adopting it, verify that screen recording and microphone permissions are granted on your OS, and read src/ to confirm how your API key and captured screen and audio data are handled.

Frequently asked questions

Official sources

1. License: GPL-3.0
2. Project website
3. README
4. Releases
5. sohzm/cheating-daddy on GitHub