Privacy

limelit is a virtual camera for macOS and Windows made by Streamfog Inc. This page says what it collects, where it goes, and how to get rid of it. The short version: your video stays on your computer; while listening is on, your audio is transcribed in the cloud and not kept; what you type to the agent, and pieces of the call transcript while a live cue is running, go through our backend to the model provider that writes the answer; and we keep an account record and usage events so we can run the service and fix it.

What stays on your computer

• Your video. limelit composites your cues over your camera picture on your computer and hands the result to the call app. No frame is ever uploaded.
• Your audio. When listening is on, limelit captures your microphone and the call's audio and streams both to Meta's transcription service (Muse Voice Transcribe), on a Mac and on Windows alike. We keep nothing of the audio. The recording and the transcript live in a session folder that is deleted when listening stops. Only the cues and your conversation with the agent survive a call.
• Your conversation log. Your conversations with the agent are a file in the app's own folder on your computer. Your cues are also kept in your cue library, below.

What reaches the model provider

The agent is a language model run by a third-party provider (currently Anthropic). Every request goes through our backend, which adds the instructions and forwards it. A request carries what the agent needs to answer:

• the messages you type and the agent's earlier replies;
• the cues on your camera, your style notes and the theme;
• files you attach, images you drop in, and screenshots of websites you ask it to read;
• during a call, pieces of the transcript. A live cue is kept current by a watcher that reads the latest transcript lines, both your side and the other participants', and sends them with its request. This only happens while listening is on and a live cue exists. Before each of those requests, the same lines, with the cue's task and its current data, go to a second provider (TypeSafe AI), whose model answers one yes/no: does anything on the cue need to change? When the answer is no, the language model is not asked.

The provider processes the request to produce the answer under API terms that exclude the content from training its models. We do not send the audio itself, ever.

What we keep

• Your account. Sign-in is with Google. We store your Google user id, email address, display name and the time of your sign-ins. The name is also used, on your computer, for the name tag cue.
• Your cue library. Your cues, their data, which are on camera and where, your style notes and the files you add to them are kept in our database and storage, so they reach every computer you sign in on and the assistants you connect. Signing in on a computer adds the cues already on it to that account's library. Kept until you delete your account.
• The waitlist. The address you give us on this site, the page it came from, and when.
• Usage events. Which features you use, when a call started and ended and in which app, how many cues were shown, whether an agent turn succeeded, app version and operating system version. Events carry your user id, never transcript text, never cue content, never a file name.
• Agent traces. During the invite-only period we keep a record of your conversations with the agent, the requests our backend forwarded and the answers that came back, so we can find and fix its mistakes. Images, PDFs and slide decks are replaced by a hash. Transcript text is never kept: where the agent read the transcript, the trace says how many lines and when, not the words. Traces are deleted after 90 days.
• Crash reports. If the app or one of the processes behind the camera crashes, a report with where it crashed, the app version, the operating system version and your user id is sent so we can fix it.
• Diagnostics, on a Mac. If the part of limelit that draws cues onto the camera stops working, or you click "Send diagnostics", the app sends a technical bundle: how that process started and ended, its log, your macOS build, and which security or device-management software is installed. It carries no cue content, transcript or file names, and is deleted after 90 days or with your account.
• Messages to Kev. What you write to Kev in the app, and his replies. Each message you send carries the app's state at that moment: app and operating system version, whether the camera and listening were on and which app used the camera, how many cues there were, and, when you flag one of the agent's replies, which reply it was and its first lines. It carries no transcript text and no cue content. Kept until you delete your account.

Assistants you connect

You can connect an assistant of your own, such as Claude, ChatGPT, Gemini or Grok, to limelit. You sign in to limelit and allow it once. From then on it can read your cue library, make cues and change them, including their data, your style notes and the files in it. It cannot reach your camera picture, your audio or your calls.

• What the assistant reads from your library and writes into it passes through that assistant's provider (Anthropic, OpenAI, Google, xAI or another) under your own account and their terms, not ours.
• We record each connection (which assistant, when), and for each thing it does, which tool it used, on which cue, how long it took and the first line of the answer, so we can fix what goes wrong. If the assistant reports something limelit could not do for you, we keep its report, which can quote what you asked. Requests that arrive without a valid connection are counted by a hash of the caller's address and browser, never the address itself.
• A connection lasts until you end it: from the assistant, in the app's settings under Connections on a Mac, or by deleting your account, which ends them all.

Who processes it

• Google (Firebase, Google Cloud): sign-in, our database, our functions and this site.
• Anthropic: the model behind the agent, as described above.
• Meta: cloud transcription while listening is on.
• TypeSafe AI: the yes/no in front of the watcher during a call, as described above.
• PostHog: usage events and crash reports.

The app also checks for updates by fetching a feed from our update server, on Windows also once a night. That request carries the app version, on Windows also what started the check, and nothing about you.

We do not sell your data, and we do not use it for advertising.

Other people on your calls

When listening is on, the other participants' words are transcribed, on your Mac or in the cloud as described above, and pieces of them reach the model providers while a live cue is running. You are the one on the call; where the law or the meeting's rules require it, tell the participants and get their consent before you turn listening on.

Your choices

• Listening is off by default and never starts without your say: you turn it on per call, or switch on listening during calls in the app's settings.
• Agent traces are part of using the agent during the invite-only period; deleting your account removes them.
• Delete your account in the app's settings, under Account. That removes your account record, your cue library, your connected assistants, your traces, your messages to Kev and your usage events. The cues and conversation log on your computer are yours to delete.
• To see, correct or export what we hold, or to leave the waitlist, write to us.

Children

limelit is not for anyone under 16, and we do not knowingly collect their data.

Changes

If this page changes in a way that matters, the app tells you at the next launch, and the date at the top moves.

Contact

Streamfog Inc. · kev@limelit.ai