Cirax
An invisible AI copilot that floats over your screen
HOW IT WORKS
What you're watching: three inputs (screen, mic, meeting audio) hitting a local Whisper, context flowing to your own model, and the answer landing in a floating glass panel that screen shares can't see.
The problem
You're in a live meeting. Someone asks a question you half-know, and the honest options are all bad: switch windows and visibly vanish, fumble through a notes file, or guess. Every "AI meeting assistant" on the market joins the call as a bot everyone can see, or sends your audio to someone else's cloud.
The question Cirax answers: can the help be present but invisible, and can it work with models you already pay for?
The approach
A small glass panel floats over everything. It combines three inputs, your screen, your microphone, and the meeting's audio (the other person's voice), and routes them to whatever model you bring: OpenAI, Anthropic, Gemini, or any OpenAI-compatible endpoint including a local Ollama.
- Transcription runs locally. A bundled Whisper instance handles speech-to-text entirely offline; a Deepgram stream is optional. The only thing that ever leaves the machine is what you send to your own LLM endpoint.
- Invisible where it matters. On Windows the panel is excluded from captures at the OS level (WDA_EXCLUDEFROMCAPTURE), so screen shares and recordings see your work, not the panel. Linux and macOS get honest best-effort modes, and the app tells you which one you're in.
- Triggers shaped like real meetings. One shortcut for "help with what's on screen," a button for "what should I say?" from live audio, follow-ups and recaps from the whole conversation, and a pill that switches to a slower, smarter model mid-meeting.
The companion problem
Answering in a panel is only half of it. If you read the answer, your eyes leave the camera and everyone knows. So a companion daemon runs a lightweight gaze-correction model (~7ms per frame, fully local) on your webcam feed and publishes the corrected feed as a virtual webcam that Zoom, Meet, and Teams pick up natively. You look at the answer; they see you looking at them.
The hard parts
- Capture exclusion is a per-platform puzzle. Windows has a real API for it, macOS 15.4+ broke the usual tricks, and Linux Wayland simply doesn't allow it. The app detects which world it's in and discloses the limit at startup instead of pretending.
- Meeting audio is not your microphone. Capturing the other side of a call means loopback capture, which is a different API on every OS, with echo and format traps.
- Secrets in Electron. API keys are encrypted through the OS keyring (safeStorage), never in config files.
The honest line
The README says this out loud: phone cameras can always see your screen, some tools can see through exclusion on older systems, and using an assistant where the rules forbid it is on you. Invisible tools deserve honest documentation.