Your Orb
Your own AI model. In your browser.
The Orb lives inside Skillerr and understands what you're working on: which tabs belong to one journey, what you and your AIs have already read, where you left off. It runs on your computer, learns from how you browse, and sends nothing to anyone. Not even us.
Measured on a MacBook Air (M1, 8 GB). Details and how to reproduce them below.
What your Orb does
Sorts your tabs into journeys
Tabs you opened together stay together; journeys on different days join when they're about the same thing. Bring over the tabs you have open in Chrome and watch them arrive sorted, with their real icons. Nothing reloads, and nothing closes in Chrome.
Finds anything by meaning
Type what you remember in the address bar ("train from Tokyo to Kyoto", "newborn feeding") and the page you had comes back, scrolled where you were.
Keeps your AIs' research
When Claude, Cursor or Codex browse through Skillerr, what they read becomes a trail beside your own: named after what they looked into, with their conclusion and sources. It's marked unfinished only from the facts: a page that couldn't open, a robot check, a step you declined.
Suggests skills from your habits
Research the same kind of thing the same way three times and the Orb offers to save how you do it as a skill your AI can follow. Worked out from your trails, with no AI model writing it.
Learns you, privately
Every week (it asks first) your Orb learns your words from your own trails and your AIs' research, on your computer. It keeps what it learned only if that files your pages better. Each source can be switched off.
Shows you everything it knows
The Your Orb page (⇧⌘Y) maps everything you and your AIs looked into around the Orb: yours in warm colours, your AIs' in cool ones. Click the Orb for its size, memory and speed, live.
Benchmarks
Everything here was measured on a MacBook Air (M1, 8 GB) with scripts anyone can run (scripts/kilr/ in the Skillerr repository). Where another model does better, it says so.
Speed and size, one text at a time (as a browser meets them)
| Per page title | Per second | Model size | Memory | |
|---|---|---|---|---|
| Your Orb (one core) | 8.4 µs | ~120,000 | 7.9 MB | ~30 MB |
| all-MiniLM-L6-v2, 4 threads | 2,171 µs | 466 | 90 MB | ~140–230 MB |
| all-MiniLM-L6-v2, 1 thread | 4,690 µs | 213 | 90 MB | ~100–120 MB |
all-MiniLM-L6-v2 run with ONNX Runtime (native code): the Orb is over 250× faster than it on 4 threads, and about 550× faster on one. On a GPU, in large batches, a transformer gets through more text per second; a browser meets titles one at a time, on your CPU.
Sorting tabs kept open for weeks into journeys
| Precision | F1 | F1 when a quarter of titles say nothing ("Log In") | |
|---|---|---|---|
| Your Orb (sittings joined by meaning) | 0.91 | 0.83 | 0.79 |
| Same rule, with all-MiniLM-L6-v2 (90 MB) | – | 0.85 | 0.81 |
| Word matching | 1.00 | 0.25 | 0.19 |
87 tabs from 14 threads of work, opened over sittings across four weeks, 20 random layouts. The Orb reaches 98% of all-MiniLM-L6-v2's accuracy on this job (F1 0.83 vs 0.85).
Real people's searches: Webis-SMC-12
8,840 searches by 127 people, split into tasks by two human annotators (Hagen et al., 2013). Deciding where one search session ends and the next begins, F1.5 as in the paper:
| F1.5 | Needs the web? | |
|---|---|---|
| Your Orb, with Skillerr's journey model | 0.922 | No, on your computer |
| Best published, without web lookups | 0.930 | No |
| Best published, with search-engine results | 0.946 | Yes |
Tuned on a quarter of the people and tested on the rest, as in the paper; our split is by person, theirs isn't published, so the numbers are close to comparable, not identical.
General sentence similarity (STS-B, SICK)
| STS-B | SICK | Size | |
|---|---|---|---|
| all-MiniLM-L6-v2 (the transformer the Orb learned from) | 82.1 | 77.1 | 90 MB |
| potion-base-8M (static, open) | 75.4 | 64.5 | 30 MB |
| Your Orb | 72.3 | 62.9 | 7.9 MB |
| Word overlap | 56.5 | 57.5 | – |
Spearman × 100 against human scores; neither test set is used to build the Orb. General sentence similarity isn't the Orb's job, and bigger models do better here.
Yours, not ours
On your computer
The model, what it learns and your trails stay in Skillerr's folder on your machine. It works offline.
Nothing sent
The Orb makes no network requests. Watch it with any network monitor while you browse.
Yours to reset
Forget what it learned, or turn it off, in Trails settings. Clear browsing data clears it all.
When you let an AI app like Claude browse through Skillerr, the pages it reads go to that AI, as they would in any browser. The Orb itself never sends anything. It chooses from the facts of your own trails and never writes text, so it can't make things up.
Only in Skillerr
Your Orb comes with Skillerr, the browser that skills your AI. Free for Mac, Windows and Linux.