Windows + Linux X11 · no macOS 14+ installer yet
AI-Sister quietly watches you work and keeps the details on your screen on this computer. When you need a phone number, an amount, a URL, or the spot where you got stuck yesterday, she answers in seconds, and every sentence opens to its evidence.
v0.1.0-alpha.171 · Windows 10+ · Ubuntu 24.04 X11 · macOS 14+: no public installer yet · Apache-2.0 · Interface in Traditional Chinese, voices in Mandarin
Where did I leave the checkout page yesterday?
You got webhook retries working. The last screen shows the tests passing; next up is the refund case.
0screenshots sent to the cloud
<100mslocal recall target
17offline characters
544bundled everyday voice lines
Not another chat window
No model rewrites the raw evidence. Her understanding can be corrected and a memory can be closed out; the screen you open is the final word.
OCR text, phone numbers, amounts, URLs, and times stay in SQLite; answers link to the local screen they came from.
AI-Sister isn't tied to one model vendor. Every typed question goes first to the CLI you picked last; it decides how to search, and AI-Sister runs the search locally.
She speaks up only with evidence, when the interruption is worth it, and within budget. The rest of the time she stays in the corner of your desktop.
Local recording has to be on. She writes what she saw, waits, and looks again.
Wait, then look
Looks are at least 400 ms apart. With no key or mouse input, she still looks again within 5 seconds.
Read the words
Windows uses system OCR, Ubuntu uses Tesseract, and macOS uses Vision.
Save the row
Words go into SQLite, and a changed picture stays only when frame storage is signed.
CLI reading has to be signed first. The CLI you chose plans the search and writes the sentences, and the search itself runs in SQLite on this computer.
You ask
The question that goes out is capped at 2 KiB.
Plan queries
The CLI returns 1 to 3 queries in everyday words.
Search here
SQLite runs them here. At most 18 sources go back, and that packet stays within 24 KiB.
Write sentences
With nothing here to cite, no sentences are asked for. One bad sentence drops the whole reply of 1 to 6 sentences.
Four CLIs can be connected. Each question goes to the one you finished setting up last.
Type a question
It reads the saved CLI again, and nothing in the program chooses a vendor.
Use that CLI
The highlight is the CLI you finished, and Claude Code is only the example.
A line she starts on her own has to pass the gatekeeper. The bar shows the points left today.
Needs evidence
Without an evidence reference, the line stays unsaid.
Clears the floor
A score under the default 0.25 stays quiet.
Points remain
A card costs 2, one line costs 1, and a glimmer costs 0.
Then she speaks
Recording is live, and it is not during the default quiet hours, 22:00 to 08:00.
Four sisters + 13 friends
17 transparent full-body workplace rigs, 17 fallback images, and 32 everyday voice lines per character all install with the app. Switching characters changes neither the facts in an answer nor the evidence.
Tap one to preview
ChatGPT · art and fixed voice lines stay on this machine
Boundaries you can see
Screens, OCR text, times, and window context stay on this machine.
Local dataSQLite stays local; the CLI can only ask for natural-language searches.
Searched locallyOnly after you explicitly agree do your question and the matching original text go to the CLI you chose.
Your own CLICharacter voices work offline; Azure read-aloud for new answers has its own separate switch.
No fallback either wayThe screen and the database stay on this computer. Each door needs its own signature.
Screen stays
Screenshots and the OCR text are written here, and no pixel is uploaded.
Rows stay
A search never sends the SQLite file.
Check signatures
CLI reading and Azure read-aloud are separate, and unreadable means unsigned.
CLI text
Once CLI reading is signed, the question and the hit text may go to that CLI.
Azure voice
It has to be signed and switched on, and only the new answer body is sent.
Claude Code + Codex
Both entry points call sister query --json on your computer and bring the matching text, time,
app, window title, and frame number back into the current conversation.
After installing, type /ai-sister-memory, then say which memory you're looking for.
~/.claude/skills/ai-sister-memory/
Claude Code skill
After installing, type $ai-sister-memory, then say which memory you're looking for.
~/.agents/skills/ai-sister-memory/
Codex skill
The skill runs only when you explicitly ask it to search your AI-Sister memory. It won't sign consents, start recording, delete anything, or operate your computer. The text and source fields a query returns enter the current agent conversation; screenshot bytes don't.
Claude Code and Codex run the same command on this computer. What comes back is text, not a screenshot.
Name the skill
Claude Code is /ai-sister-memory. Codex is $ai-sister-memory.
Run it here
It runs sister query on this computer and asks for 10 rows first. It asks again only if the missing rows could change the answer.
Text comes back
What comes back is the time, app, window title, and a frame or chunk id.
Three steps
Setup for Windows, the .deb for Ubuntu 24.04 X11. macOS has no public installer yet. The app itself is in Traditional Chinese.
Turn on local recording first; decide on frame storage and CLI reading separately.
Ask what you were doing yesterday. Every fact comes with its source.
Setup includes the desktop app, the CLI, and the WebView2 offline installer.
The .deb installs the desktop app, the CLI, AT-SPI, and local Tesseract.
The native build joins a release only after it passes Developer ID notarization and on-device permission and memory acceptance tests.