The companion

Homestead Companion lets your own Firefox fetch updates from the people you follow. When you visit a followed profile, its new posts are saved for the household. It has three parts.

Part	Runs on	Job	Status
Companion extension	Firefox 140+ on your computer	Waits for a followed profile to load, takes the page and the list of images it used, takes a screenshot, and sends them to Homestead. It knows nothing about any one site.	Built on `main`; store listing Coming in 1.0 (E-04)
Homestead Helper	Your computer	A small program Firefox starts. It reads the images from Firefox’s own disk cache and uploads them, so nothing is downloaded twice.	Built on `main`; release downloads Coming in 1.0 (E-04)
Server extraction	Your Homestead	Stores the captures, then asks an AI model (through OpenRouter) to point out each post, its link and its images. Writes short summaries.	Available on main (S-16) Coming in 1.0 (S-17)

Until the server side lands, the extension and helper have nowhere to send captures: the captures API answers 501 not_implemented on main.

Two ways it saves posts

Passive capture. You open a followed profile, story or post on a platform you turned on. The extension waits for the page to settle and captures it, and records a visit. The toolbar icon shows a honey dot for 3 seconds. Nothing else changes on the page.

Update circle. A button on a circle page opens that circle’s profiles for you, 3 tabs at a time, in a collapsed tab group. Each tab loads, scrolls down up to two screens, gets focus for about 1.5 seconds so images render, is captured, and closes itself. A control tab shows progress and ends with something like “Family is up to date · 12 new updates”, and a button back to Homestead. A circle holds at most 30 capturable accounts, so a run is at most 30 tabs.

Available on main

The extension never clicks or types on a site. Stories that need a tap to open (Instagram, for example) are saved when you watch them yourself, not during an update.

What it reads, and what it doesn’t

It captures only exact matches for people you follow, on platforms you turned on. It never asks for access to all sites, and never runs in private windows.
Before sending a page it removes scripts, styles, form field values and editable text.
The helper only reads Firefox’s cache2/entries folder, never cookies, passwords or history, and never writes to your Firefox profile.
The helper only uploads to the Homestead address you gave it at install, only images and videos from the platform’s media hosts, with a token that’s valid for one capture for 10 minutes. It never fetches anything from the network.
If an image isn’t in the cache, it’s recorded as missing, not downloaded.

The AI part

Coming in 1.0

AI extraction is off until the owner adds an OpenRouter API key. Then each capture goes to an image-capable model (default xiaomi/mimo-v2.6-pro) with the page text, page images and a screenshot. Routing only uses providers that don’t collect data.

The model can only point: it answers with IDs the server gave to links, images, text lines and times on the page, plus a few fixed values. It can’t add a link, a script or any text of its own. Every link, image and caption that gets stored comes from the server’s own reading of the page. The model can still be wrong, for example by picking the wrong image for a post.

Optional short summaries (up to 280 characters, links stripped, labelled “AI”) are the only free text a model writes. Each member can hide them. Available on main (S-18)

Raw captures are kept until a disk quota (20 GB by default), then deleted oldest first, so the owner can re-run extraction after changing the prompts or model.

Recently visited