Skim Recap reads page text only to work out what you skipped and to summarise it, and that work happens on your own machine. This page lists exactly what is stored and what is transmitted.
01 // The boundary
Where the summarising happens
Summaries run locally with Gemma 4 E4B through LiteRT-LM inside
Chrome. The model downloads once from Hugging Face and is cached
by the browser for later sessions.
Network access is used for model retrieval, while skipped article
text is not sent to a hosted LLM API for summarization.
The one-time download is 2.97 GB, and it is the honest cost of the last column being zero.
02 // What leaves the machine
One request, on first use
Data
Where it goes
Article text you scrolled past
Stays in the browser
The recap that is generated
Stays in the browser
Scroll activity and timing
Stays in the browser
Your settings
chrome.storage, on your profile
The model file, once
Downloaded from Hugging Face on first use
There is no account, no server of mine, and no telemetry. After the
model has been cached, the extension works with the network off.
03 // Permissions
Why each one is requested
offscreen
Hosts the WebGPU inference runtime — the only Manifest V3 context with GPU access. No visible interface.
