Silent input // ocuclaw
←  Main page
OcuClaw 2.1 · Silent input

Type a message. Without saying a word.

For meetings, quiet trains and sleeping babies. Silent input types on your Even G2 glasses using just the R1 ring. One swipe or tap per letter, and it guesses the word. Real chat messages average about 3.6 moves a word.

Even Realities links are affiliate links · OcuClaw may earn a commission from eligible purchases.

A friend asks if you’re still coming to dinner. The reply takes 14 words and 20 moves, and most of the words need no letters at all.

R1 ring or glasses temple

A message comes in. Tap then hold opens the menu, and Silent input is right at the top.

  1. A message comes in. Tap then hold opens the menu, and Silent input is right at the top.
  2. Down, up, then a hold for “Yes”. The comma comes by itself.
  3. “I” and “am” are already on top of the next-word list: one hold each.
  4. Most words need no letters. The hold just runs longer when the word sits lower.
  5. “soon” takes a swipe and a tap. Fourteen words, twenty moves.
  6. The page menu opens on Review draft. The full stop is added for you, and one tap sends it.

Recorded in the simulator. The R1 ring shows each press and what it did. The glasses' temples work too.

01 Try it

Three groups of letters. Then the word.

The alphabet is split into three groups. For each letter you just say which group it's in, and the list narrows to words that start that way. Your word often shows up after two or three letters. Hold to step down the list (one word every 400 ms) and let go to pick. Before you start a word, a quick hold opens Tools instead: comma, full stop, question mark, and Review to send.

Swipe up
↑ A–I
The first nine letters.
Tap
· J–R
The middle nine.
Swipe down
↓ S–Z
The last eight.
Hold
Pick
Hold to step down the list, let go to pick. A quick hold between words opens Tools. Double tap to undo.
· J–Rtap
Words 0Moves 0Per word –Learned 0

Loading the word list…

Click the top, middle or bottom of the ring, or drag up and down on it. Double tap the middle to go back. Try “I will be there in five minutes”: 11 moves for 7 words. Send it and type it again: 7. Keyboard: ↑ → ↓, hold Enter, Backspace.

This is the real word list from OcuClaw 2.1 (30,051 words), ranked by the same rules as the phone. It matches the app on all 1,380 lists we checked. It learns from you for this visit only, and nothing leaves the page. Missing here: My words, Jev and the exact-match page. A few slurs are hidden. Word data: Google Books Ngram (CC BY 3.0), word lists by hackerb9/gwordlist (CC BY 3.0), Mozilla Common Voice sentences (CC0), Tatoeba sentences (CC BY 2.0 FR, tatoeba.org), Taskmaster dialogs by Google (CC BY 4.0), OpenAssistant oasst1 conversations (Apache 2.0), AMI Meeting Corpus transcripts (CC BY 4.0), AAC communications and language model by Keith Vertanen and Per Ola Kristensson (CC BY 4.0, aactext.org).

Word not on the list? Spell exactly shows every match from 177,113 more words, including names, places and tech terms. Still not there? Spell it letter by letter (three moves a letter), and it's remembered next time.

02 The list

Your phone does the guessing, offline.

Every move re-ranks the list on the phone itself, so there's no network wait. It looks at the last two words, the message you're replying to, and how you usually write. Optional outside help can reorder the list, but never add words to it.

fig.01 · one move, one listmove · match · rank · pick · learn
flowchart LR
  classDef you fill:#0e1622,stroke:#7ad3ff,color:#dbeeff,stroke-width:1px;
  classDef cc fill:#0c1a10,stroke:#4dd58a,color:#dff5e6,stroke-width:1.2px;
  classDef box fill:#1a1408,stroke:#ffb454,color:#ffe7c2,stroke-width:1px;
  classDef dim fill:#0f1310,stroke:#4a5048,color:#cfe8d4,stroke-width:1px;
  M["ring move<br/>↑ A–I · tap J–R · ↓ S–Z"]:::you
  subgraph P[" on your phone "]
    direction LR
    F["match<br/>words that start<br/>with these groups"]:::cc
    W["word model<br/>30,000 words<br/>400,000 pairs<br/>300,000 triples"]:::dim
    C["context<br/>the last two words<br/>the message you answer"]:::dim
    L["what it learned<br/>from every word you pick"]:::dim
    R["rank"]:::cc
  end
  J["Jev by TypeSafe<br/>optional · on your host"]:::box
  G["top words<br/>on the glasses"]:::cc
  H["hold, let go<br/>the word goes in"]:::you
  M --> F --> R --> G --> H
  W --> R
  C --> R
  L --> R
  J -.->|"reorders"| R
01
It learns how you write
It remembers your recent words, the pairs you use (“on my way”) and words you had to spell out. Undo forgets what it learned. It all stays on your phone, and you can reset it.
02
Optional help from your host
Add a TypeSafe key and your host asks their Jev model which word is likely next. It only reorders, and saves about 6% of moves. Your agent can also suggest up to three replies.
03
What leaves the phone
With Jev off, nothing until you hit send. With Jev on, the end of your draft and the message you're answering go to TypeSafe. What it learned about you never leaves.

We tried having a chat model write the next words. Each call took 2.6 to 6 seconds, and overall it was no faster than the phone on its own (3,096 ms a word against 3,095 ms). Picking from known words won.

03 How it was tuned

One number to beat: moves per word.

A replay tool types thousands of real chat messages through an exact copy of the keyboard and counts the moves. Each change is tuned on half the messages, then tested once on the other half.

Moves per word · held-out chat · Sept 2026Lower is better
Where we startedfrozen baseline
6.62
Contractions, reply words, exact-match pagethree changes
5.04
Tables built from chat30,000 words · 400,000 pairs
4.41
Three-word context, reply tuning, pairsfour changes
4.31
Learning spelled wordson the phone
3.85
More words to match exactlyshipped in 2.1
3.63

Turn on next-word suggestions and it drops to 3.20 moves a word.

When you can't say it, you can still send it.

What you need: smart glasses (the Even G2) with the R1 ring, the OcuClaw app on your phone, and OcuClaw 2.1 on your OpenClaw or Hermes Agent host. Jev is optional.

Set it up →