VibeSearchMaya Okafor @mayabuilds
Spent a weekend rebuilding my bookmarks as a tiny search engine. The 9 tools I actually use, with prices.
- Free
- No account
- Works offline
Your browser’s memory, finally searchable.
Right-click to save anything. A search engine that lives inside your browser finds it by meaning, reads the text in your screenshots, and keeps working with the wifi off.
- Chrome
- Brave
- Firefox
/
Found by meaning The video never says “thing”. It was found by what it means.
Best matchSalt the water like the sea: the 4 minute pasta lesson
…salt the water till it tastes like the sea…
Why you should save your pasta water
The only birthday cake recipe you need
Example library · 4,218 saves Offline. Nothing left this computer.
Right-click anything. It saves the right thing.
Not the page you were on: the post, the video, the quote, the image. On 18 sites it knows exactly what you pointed at, even in the middle of a feed.
Dev Patel @devpatel
Three CSS tricks I wish I had known sooner. The third one deleted 200 lines of my stylesheet.
Lena Hoffmann @lenaruns
Sunrise 10k along the river. Slowest pace of the year, best run of the year.
Eighteen sites it knows by name.
Posts, videos, repos, threads and products come back with the right author, words and media. On every other page it reads the page’s own metadata, so you still get the real title and image.
- X
- YouTube
- GitHub
- TikTok
- Medium
- SSubstack
- Stack Overflow
- YHacker News
- Spotify
- Vimeo
- Twitch
- Product Hunt
- Notion
- Figma
Save without reaching for the mouse.
- Quick-save this tab
- Screenshot this tab
- Open the popup
Command on a Mac. Rebind any of them in your browser’s shortcut settings. A small note on the page confirms each save, with Undo.
Save the whole window in one click.
Every open tab goes into a folder you name, ready to search. Or paste a link, grab a screenshot, or archive the page from the same popup.
Drag a box around anything to save it.
Capture any part of any page. The text inside the box is read and becomes searchable.
Archive the page itself.
A full, self-contained copy. It still opens after the site changes or disappears.
File it while you save it.
The right-click menu lists your spaces and folders, with the ones you use most on top.
Every keystroke runs four searches at once.
Most search boxes match letters. This one also matches meaning, typos and sound, all at once, then weighs them into a single ranking. Meaning counts most, so a lucky keyword never beats the thing you actually meant.
- MeaningAn AI model on your computer compares what you meant with what each save is about.…
- Exact wordsEvery word, weighted by where it sits. A hit in the title counts more than one in the body.…
- Typos and half wordsMisspellings and words you stopped typing halfway still land.…
- Sounds likeTitles and names are matched by how they sound, not how they are spelled.…
Relevance0.7 × meaning 0.38 + 0.3 × words 0.92 = 0.00
Top result
Kubernetes explained in 100 seconds
And when you know exactly what you want, say it.
A small search language for power users. Type a filter and it turns into a chip you can read, edit or remove. It suggests fields and values as you type, so there is nothing to memorise. Put a minus in front of anything to leave it out.
source:reddit- Only Reddit
added:last30d- Saved this month
has:video- Only saves with video
likes:>=1000- A thousand likes or more
-tag:archived- Leave out a tag
(pytorch OR jax)- Either word
sort:updatedAt desc- Newest change first
/private- Search your locked spaces
It reads your screenshots, right on your computer.
Screenshots, photos of whiteboards, receipts, slides. VibeSearch finds every line of text inside them and makes it searchable, by the exact words and by what they mean. Nothing is uploaded to be read.
Saved. Waiting its turn.
- Directions for Saturday!98%
- Door code is 4471#97%
- Flat 3B, second floor96%
- Park behind the bakery93%
- Wifi: casa-augusta91%
door code1 match · Screenshot · saved 4 days ago
A real text reader, built in.
PaddleOCR’s PP-OCRv6 runs two models: one finds the lines, one reads them. Both run in WebAssembly, downloaded once and cached for good.
Walled off from everything else.
The reader runs in its own sealed sandbox page. It gets the picture, hands back text, and touches nothing else.
Searchable by meaning too.
Text it reads is folded into the save and the save is re-understood, so “where do I park” finds the chat about the bakery.
Copy text out of any image.
Right-click a picture on any page and pick Extract text from image. Low-confidence lines are dropped, so you get the words, not the noise.
A whole search engine, running in a browser tab.
Search companies run this in data centers: an AI model, a vector database, a keyword index and a text reader. VibeSearch fits all of it inside your browser, and none of it needs a server. Here is one search, start to finish.
cooling data centers site:arxiv.org
17.7milliseconds
Start to finish, on your own computer. No request left the building.
Seawater cooling for hyperscale data centresarxiv.org · relevance 0.83
Typical timings on a laptop, for an example library of 4,218 saves. The very first search after the browser starts also loads the model.
Total, about 18 milliseconds.
What’s running, all of it on your machine.
The models download once, the first time they’re needed, and stay cached. After that, everything here works with the wifi off.
- Embedding model
- mdbr-leaf-irBuilt for search. 768 dimensions per save
- Text reader
- PaddleOCR PP-OCRv6One model finds lines, one reads them
- Model runtime
- ONNX RuntimeWebAssembly with SIMD, inside the browser
- Vector store
- Our ownOne packed binary file in private browser storage
- Keyword index
- FlexSearchExact, fuzzy, half-word and sounds-like
- Database
- RxDB on IndexedDBReactive, so every open view updates itself
Four rooms, each with one job.
A browser extension lives under strict rules: the background can be shut down at any second, and code isn’t allowed to do whatever it likes. So VibeSearch is split into four parts that each work within those rules, and four helpers that keep heavy work off the screen you’re using.
What you see
Search page and popup
Never touch your data directly. They ask, and render the answer.
Query worker · parses every keystrokeThe switchboard
Background worker
Owns menus, shortcuts and screenshots, and passes requests to the brain. Holds no data, because the browser can stop it whenever it likes.
The brain
Offscreen document
Database, keyword index, vector store and every background queue. Stays alive while the switchboard comes and goes.
- Embedding worker · runs the model
- Ranking worker · fuses and sorts
- Compaction worker · tidies the vector file
The locks between rooms
Every message is checked, twice
- Is it really us? Messages from web pages pretending to be the extension are turned away.
- May this page ask? Unlocking a private space only works from the search page itself.
- Is it on the list? The brain answers only the requests it was built for. Some it takes only from itself.
The reader
Sealed sandbox
The text reader needs rules the rest of the extension is never given, so it gets its own walled-off page and only ever sees a picture.
Still fast at fifty thousand saves.
The engineering under the search box, one piece at a time.
It brings its own vector database.
Every save becomes 768 numbers, about 3 KB, packed into one file in your browser’s private storage. New saves are added to the end and never rewritten. Searching scans it straight from memory: exact results, nothing approximated, and no index to build or corrupt.
56 49 42 45Magic“VIBE”01 00 00 00Version100 03 00 00Dimensions7687A 10 00 00Saves4,218Similarity one multiply-add per number. No square roots, because every vector is already unit length.
Import ten thousand bookmarks without a crash.
AI models slow down sharply when long and short texts share a batch, because the short ones get padded to the longest. VibeSearch sorts every import by length and feeds the model small, even batches, so a huge import stays light on memory.
Your search cuts the line.
Imports and searches share one model with two queues. Yours always goes first, so search stays instant while five thousand bookmarks are still being read.
Pull the plug halfway through and nothing breaks.
Every few hours the vector file is rebuilt to drop what you deleted. Before touching anything it writes down what it is about to do. If the browser dies mid-way, the next start reads that note and rolls back, or rebuilds from scratch. It never shows results from a half-finished file.
- Idle
- Preparing
- Prepared
- Swapped Power cut here? Restored on next start.
- Committed
It keeps working after you close the tab.
Four background jobs carry every save from saved to searchable. Each one remembers where it stopped, retries what failed with growing pauses, and picks up again after a browser restart.
- SavedInstantly, even offline
- Details fetchedTitle, image, author
- Text readInside every picture
- UnderstoodTurned into meaning
- SearchableBy word and by meaning
Big imports don’t flicker.
Changes are gathered for a fraction of a second and sent as one update, so the page redraws once instead of five thousand times.
Bring the bookmarks you already have.
Your whole bookmark tree and every GitHub star come in with one click each. Titles and previews fill themselves in, and everything is searchable the moment it’s read.
Browser bookmarks
- Bookmarks bar 214
- Recipes 61
- Design refs 148
- Other bookmarks 1,002
- Reading 386
- Trips 74
- Mobile bookmarks 68
1,284 bookmarks, every folder kept, all searchable
Import again any time. It updates what changed instead of making copies.
GitHub stars
ollama/ollama142ktldraw/tldraw38.1kexcalidraw/excalidraw97.4kpmndrs/zustand52.6k
249 starred repos, with their star counts
Search likes:>=10000 to find the big ones. Unstar a repo and it leaves your library on the next import.
Built for thousands of saves, not thirty.
- Spaces, folders, tab groups
Nest them as deep as you like. A library of thousands still makes sense.
- Tags with colours
Rename, recolour and favourite them. Filter by any of them in a word.
- Organize mode
Press O and drag saves between folders and spaces.
- A 30-day bin
Deleted something by mistake? It waits a month before it’s gone.
Share a whole folder as one link.
Pick a folder, a space or a handful of saves. Your friend opens a page they can browse, or imports the lot into their own VibeSearch with one click.
- The secret lives only in the link.
- Our server keeps a fingerprint, never the key. Without the link, nobody gets in.
- Add a PIN if it’s private.
- Then only the people you tell can open it or import it.
- Hidden from Google by default.
- Search engines stay out unless you switch it on.
- Switch it off whenever you like.
- Revoke a link and it stops working for everyone, everywhere.
Your library lives on your computer.
No account, no server holding your saves, no one reading your searches. The search engine runs where your data already is.
Stays on your computer
- Every save and folder
- Every search you type
- The AI models
- 768 numbers per save
Turn the wifi off. All of this keeps working.
Leaves only to do a job
- A link’s address, to fetch its title and picture, the way a chat app previews a link.
- Saved images, backed up so clearing your browser never loses them.
- Shares and Drive backups, only when you ask.
Lock the spaces that matter.
Put a password on any space. It locks itself after five quiet minutes and stays out of every search until you open it again. Forget the password and your recovery questions get you back in.
The password itself is never stored, only a salted fingerprint stretched 180,000 times.
Take everything out, any time.
Back up to a file, to your own Google Drive, or to a Google Sheet you can read. Restoring merges with what’s there, so running it twice never makes duplicates.
- File
- Drive
- Sheets
Ten seconds to install, and free for good.
- 1Add it
One click from your browser’s store.
- 2Pin it
Keep the V next to your address bar.
- 3Press Ctrl Shift S
On any page. That’s your first save.
Before you install.
Is it really free?
Yes. No account, no trial, no cap on how much you save. VibeSearch, the paid app, adds your phone and more, but the extension never needs it.
Does it work offline?
Saving, searching and reading images all work offline once the models have downloaded. A link’s title and preview fill in when you’re back online.
Which browsers?
Chrome, Brave and Firefox 142 or newer. Reading text inside images needs Chrome or Brave. Everything else works in all three.
Will it slow my browser down?
No. The heavy work runs on background threads in small batches, and your searches always go first. You won’t feel it running.
How big is it?
The extension is small. The first time it needs them, it downloads its AI models once and keeps them cached, so it never downloads them again.
Can I bring my bookmarks?
Yes, your whole bookmark tree and your GitHub stars, one click each. Folders come across as they are, and importing again updates instead of duplicating.
Want it on your phone too?
VibeSearch takes it everywhere.
The full app adds your phone, search inside videos, notes that sync, and one-click imports from X, Reddit and Instagram. One library, on every device.
$39 to start, then $7.50 a month.
See VibeSearch