Pavo20 camera mount and antenna holder
A printable mount that moves the camera forward and protects the antenna on a compact FPV build.
Unbury searches your browser bookmarks by meaning, not just matching words. Ask for “that 3D-printed part for the drone” and find the page even when its title says something else.
Pavo20 camera mount and antenna holder
A printable mount that moves the camera forward and protects the antenna on a compact FPV build.
Whoop accessories collection
Replacement ducts, mounts and printable parts for small cinewhoops, including the Pavo series.
Building a lighter Pavo20 Pro
A practical build log covering mounts, screws and weight-saving modifications.
A folder tree answers "where did I file it". These are the questions people actually have, and they are the ones Unbury was built for.
“What did I save this month about AI agents?”
A subject plus a stretch of time. The time phrase is pulled out and used as a filter, so it narrows the dates instead of muddying the meaning.
“That 3D-printed part for the Pavo 20”
You remember what the link was, not what it was called. Search by meaning is the only kind that survives that.
“Did I already have something like this?”
Before you save the fourth article on the same thing, ask whether the pile already has it.
A right answer scores about 0.60 here; unrelated links sit at 0.44 to 0.47. That gap is narrow, so Unbury keeps only what is within 12% of the best result, and only when the best clears the bar. When nothing does, the screen says so and offers the weak matches on request instead of dressing three of them up as answers.
Nothing fires while you type. Half a sentence is a different question when you search by meaning, so search happens on Return.
Change the question and the match count and score disappear, and the list beneath is labelled with the question it actually answers.
The saved page is close enough to trust. Weak matches stay out of the way.
Ask Unbury is not handed a list of results. It runs the search itself, and when the first attempt comes back weak it rephrases and goes again. Every search it ran and everything that came back stays in the transcript, and the evidence panel separates the links that carry the answer from the ones it read and set aside.
It never changes a link. It does keep the conversation — an index beside your library, with a sidebar, new conversations and delete.
Two of the three engines are command-line tools already on your Mac — Claude Code and Codex — and they spend the subscription you have rather than your API key. The third sends the work to a model on OpenRouter, which is far quicker and costs fractions of a cent. The list only offers the CLIs it can actually find installed.
Your links live in ~/Library/Application Support/Unbury: vault.json, which is plain text a person can open, and vectors.bin, a flat block of numbers. Around 3 MB for 600 links, with your settings and any saved conversations in the same folder. No account, no server, no sync you did not ask for.
With nothing typed, the search screen is the tags your collection is made of — the most-used first, each with the number of links behind it. Browsing your own collection never costs anything.
Pick several tags at once and the cloud collapses to what is still reachable from where you are.
How many links, from how many sites, over how long — then the months you saved in, the sites you saved from and the words you file under, ranked against each other. Click any bar to see what it is made of.
794 links from 268 sites, over 5 years.
Click any bar to see what it is made of.
is all it took to collect half of the 794 links here. The other half took 39.
is the busiest month here — 5× a typical one. Click that bar to see what it was.
of everything comes from 10 sites. Another 214 sites you visited once and never again.
is the longest stretch with nothing saved at all, from Apr 2022 to Aug 2022.
| Year | Links saved | Busiest month | Months with nothing saved |
|---|---|---|---|
| 2022 | 50 | Nov 2022 · 11 | 5 |
| 2023 | 126 | Nov 2023 · 15 | 0 |
| 2024 | 168 | Nov 2024 · 19 | 0 |
| 2025 | 222 | Sep 2025 · 26 | 0 |
| 2026 | 228 | Mar 2026 · 61 | 0 |
268 distinct sites
214 sites appear exactly once.
187 tags · 2.5 per link
A handful cover most of it; the rest appear once or twice.
You save a link while browsing, then import when you feel like it. There is no watcher and no scheduler, on purpose. Unbury reads your browser profiles and never writes to them.
Which profiles it found, how many links are new, and roughly what the import will cost — before it starts, not after it has spent your money.
In batches, so stopping halfway never throws away work already paid for.
The link stays, marked as no longer in the browser, and still turns up in searches. Losing something because a browser was tidied is worse than keeping a stale row.
18 links · about $0.01, and under a minute
read and described once each, from your own key. Browsing them afterwards never leaves this Mac.
Unbury costs nothing to install and is not a subscription, but two of the things it does spend money — through an API key you supply, on models chosen for being cheap and good rather than famous. OpenRouter by default; OpenAI if you would rather not have a second company in the middle.
The link is fetched and stripped to text. Sites that refuse to be fetched still get described, from the title and the domain.
One line in plain English, plus tags — reusing a tag you already have before inventing a variant of it.
A vector, which is what makes "the drone part" find a page titled something else entirely.
Ranking 600 links takes about a millisecond, locally. When you search, the only thing that leaves the Mac is the question itself, to be turned into numbers.
Nothing leaves your Mac.
The question is sent to be vectorised. Real, but thousandths of a cent.
The two that cost. Both say what they will cost before they run.
The app costs nothing and the source is MIT-licensed, but it is not free to run: it uses an API key you pay for, from OpenRouter or OpenAI. Importing 600 links cost about $0.42. A search costs thousandths of a cent. Browsing and filtering your library cost nothing at all.
No. Your library stays on your Mac. When you press Return on a search, the question you typed is sent to the provider whose key you supplied, to be turned into numbers — the question, never the library. The privacy page names every moment something leaves, including the two you would not think of.
Brave, Chrome, Arc, Edge and Vivaldi. Safari and Firefox keep their bookmarks in a different format and are not read yet. Unbury reads those bookmark files on your Mac profile by profile, shows you which profiles it found before importing anything, and never writes to them.
A Mac with Apple Silicon running macOS 14 or later, and an API key you paste into Settings — OpenRouter by default, or OpenAI. Keys are kept in the macOS Keychain, not in a file beside your data.
Because a search tool that always returns something cannot be trusted. When no link is close enough to the question, Unbury tells you the best score it saw and offers the weak matches on request.
Yes. Unbury checks for new releases itself and has since its first release, and you can also check by hand from the Unbury menu or from Settings.
Signed with a Developer ID and notarised by Apple, so it opens by dragging it to Applications — no right-click-open dance, no "damaged" warning.