Privacy

Nothing you type leaves your device

Obadh does not send what you type to a server. Typing and suggestions run on your device, without an account, analytics or typing telemetry. What Obadh learns stays there, and you can erase it. The platform sections below explain what each released app sends and keeps.

A keyboard handles your messages, searches, and the words you would not say out loud. Suspicion of one is reasonable, and a privacy page is not evidence. Checking is. The sections below describe the behavior and name the files behind it.

This page describes the released apps: Obadh 1.0.1 for iPhone and iPad, and Obadh 1.1 for Mac. Linux, Android, Windows and ChromeOS are planned; their privacy details will be described here before they are released.

It collects nothing about you

There is no account, no sign-in, no email address, no advertising identifier, no analytics SDK, no crash reporter, and no telemetry of any kind. Obadh does not know how many people use it, which is a cost the project accepts.

Typing and suggestions stay on your device

Transliteration, autocorrect, next-word suggestions where available, and emoji search run locally against data bundled with the app. They do not send the text you type to a server, and typing still works without an internet connection.

Both apps use the Rust engine. Its runtime dependencies provide serialization, finite-state transducers, memory mapping and hashing; the engine contains no HTTP client or socket code.

On iPhone and iPad, neither the app nor the keyboard connects

The released Obadh app and keyboard extension for iPhone and iPad contain no networking code: no URLSession, no HTTP client and no socket calls. Neither makes update checks, downloads models or opens a network connection. Granting Full Access does not enable a network feature in either of them. Apple’s App Store handles app downloads and updates separately.

The iPhone and iPad download includes the autocorrect lexicon and next-word model, so suggestions work in airplane mode too. The released app’s App Store privacy label says Data Not Collected.

On a Mac, a separate updater connects

Obadh for Mac’s input method contains no networking code. A separate app inside it, Obadh Updater, checks GitHub for releases. Automatic checks are on by default and start when you activate Obadh. After a successful check, another automatic check is not due for 24 hours. Failed checks may be retried after an hour; opening the updater or choosing Check Now can check sooner. Turn automatic checks off in Obadh Settings, under Updates. Manual checks and update downloads remain available.

Checks go to GitHub’s release API. If you choose Update Now, the updater also downloads the selected release’s package from GitHub, following HTTPS redirects to its release CDN on subdomains of githubusercontent.com. Neither request sends your typing, learned words, cookies, account credentials, or an app-generated user or device identifier. The fixed User-Agent is Obadh-Updater, with no installed version or system details in that header.

GitHub and the servers delivering its assets can see your source IP address, the time of the request, the requested path and HTTP headers. The package download’s path identifies the release being requested. GitHub handles this request information under its privacy statement. Opening a release page in your browser uses that browser’s own cookies and settings, separately from the updater.

The updater runs separately from the input method and does not receive keystrokes. It installs an update only after you choose it, macOS authorizes administrator access, and the package passes the developer-signature and Apple-notarization checks. macOS may also contact Apple for its own security checks; Apple describes those checks separately from GitHub’s update traffic.

What the keyboards read while you type

Each keyboard uses the input needed to compose the current word. Obadh does not keep an app-by-app typing history; the local learning that remains is described in the platform sections below.

On iPhone and iPad, editing a word around the cursor and offering next-word suggestions also use nearby text. That processing stays on the device. iOS does not allow a custom keyboard into a secure text field: tap a password field and the system keyboard takes over.

Everything it keeps on your iPhone, in one list

The personal suggestion snapshot is a compact binary the engine exports, holding what it has learned about which words tend to follow which: counts over short runs (up to three words of context and the word that came next), plus the spelling of any word the shipped word list has no token for. It sits in the container the app and the keyboard share, at Library/Application Support/ObadhKeyboard/personal-autosuggest.snapshot. It is bounded in size, and it is validated on load against a fingerprint of the shipped model, so a snapshot built from a different generation of the data is discarded instead of imported.

The learned-word store holds words you have typed that the built-in lexicon does not know (names, slang, brands) and any word at all, known or not, whose correction you refused by tapping your own spelling. It is there to stop autocorrect fixing a word you have shown it you mean. One thing it does not hold back on iPhone and iPad: a word that matches the English loanword list exactly is still written in Bangla.

It does not believe you the first time. Each word accumulates evidence and is protected only once that evidence crosses a threshold, so a one-off typo cannot immunize itself. Refusing a correction protects the word at once; ordinary use takes three commits. Evidence halves after thirty days without use, so a word you have stopped typing fades out on its own. The store lives in the shared preferences under the key keyboard.learnedWords, holds at most 500 words, and keeps a score and a timestamp for each one.

Your settings sit in the same shared preferences: typing sounds and haptics, each on its own switch, which language emoji search opens in, whether corrections may be inserted automatically, and when the keyboard last confirmed it had Full Access. The emoji panel remembers which skin tone you picked for a given emoji, and up to 32 recents, each with a timestamp and a use count that fades, which is how it decides what to drop.

These stores are not a running transcript or a history of what you typed in each app. The personal snapshot can retain short word sequences as described above; the timestamped entries are the learned-word store, emoji recents and Full Access status.

On a Mac it receives the keystrokes you type through it

While Obadh is the active input method in a text field, it receives the keystrokes needed to compose the text. Its input method runs outside the App Sandbox, because Text Input Services will not register a sandboxed input source. What differs between one input method and the next is what happens after the keystroke, and here that is a transliteration, a lookup in a local lexicon, and forgetting.

Settings and learned words on your Mac

The macOS input method stores its settings in its own preferences domain, org.unmukto.inputmethod.obadh: whether the candidate bar is shown, how many suggestions it offers, whether auto-insert is on, and whether to check for updates. Four more entries record the last successful check or completed update, the last automatic-check attempt, the version currently offered when it is newer, and the last version announced in a notification. These stay on the Mac.

The learned-word store uses the key learnedWords in the same domain. Like the iPhone and iPad store, it keeps words, scores and timestamps to stop unwanted corrections. Auto-insert is off by default on both platforms. The Mac download includes the spelling lexicon, but no next-word model or personal next-word snapshot.

macOS also keeps update notifications and local diagnostic logs. The updater logs versions, outcomes and error details, not keystrokes; Obadh does not upload those logs.

For testing, an administrator can create /Library/Preferences/org.unmukto.inputmethod.obadh.updater.plist with a different update-feed URL. Check and download requests then go to that server instead of GitHub, with the same package-verification requirements. This file is not created by a normal installation and contains no typing data. The update documentation describes how to set and remove it.

Erase what it has learned

On iPhone and iPad, open the Obadh app, go to Privacy, and choose Clear Learned Words. That deletes the personal snapshot and empties the learned-word store in one action. Suggestions from the built-in lexicon are unaffected, because they were never yours in the first place.

Two things it does not reach. A keyboard still running holds its own copy of the snapshot and writes it back after the next word you commit, so switch away from the Obadh keyboard before you clear. And it leaves the emoji recents alone. Deleting the app removes its container, which takes those and everything else above with it.

On a Mac, open Settings from the input source’s menu in the menu bar. Under Autocorrect a row says how many words are currently protected from correction, with a Reset button beside it that empties the whole store. That row is the only place the learning is visible or undoable, which is why the count is shown at all. It counts protected words only, and the button is greyed out when that count is zero, so a store holding nothing but words below the threshold has to be cleared from a terminal: defaults delete org.unmukto.inputmethod.obadh learnedWords does it.

Full Access gets you sounds, haptics and your own settings

iOS asks you to allow Full Access before a keyboard extension may do two things Obadh needs: play key sounds and haptic feedback, and read and write the container it shares with its own app. The switch is step three of turning the keyboard on, and it can be turned back off in the same place.

Without it the keyboard still types. Transliteration, autocorrect and suggestions all work, because the data they need is bundled inside the extension. What you lose is the key sounds and the haptics, and everything in the shared container the extension cannot reach without it: the settings you chose in the app, and the words the keyboard would otherwise have learned to leave alone.

Being wary of the prompt is fair, because Full Access is exactly the permission a keyboard would need in order to collect what you type and ship it somewhere. What it does not do is give Obadh something to send: the keyboard’s entitlements file lists exactly one entitlement, which is the shared container, and neither target contains code that opens a connection. Do not take a toggle’s word for that. Checking it takes about a minute, and the next section says how.

Check every line of this yourself

Clone obadh-ios and check out the released v1.0.1 tag, then search the app, keyboard and shared Swift sources for URLSession, URLRequest or socket calls: there are none. URLs in plist DTD declarations, comments and developer scripts are not network calls in the app. Scripts that fetch Unicode, CLDR or lexicon data run on a developer’s machine and are not compiled into the app.

In that iOS tag, Config/ObadhKeyboard.entitlements lists one entitlement: the app group. The seventy-three lines of Shared/Sources/Engine/PersonalAutosuggestStore.swift handle the personal snapshot file. Shared/Sources/Settings/LearnedWordStore.swift holds the learned-word storage format: a dictionary of words, each with a score and timestamp.

For Obadh for Mac 1.1, the networking code in obadh-macos is in Sources/Updater, which the input method does not include. The release script, scripts/package.sh, checks the input method and installer binaries for known networking symbols, including URLSession and socket APIs, and rejects a build if it finds them. docs/updates.md describes the updater’s requests and verification. If you find something this page does not describe, reporting it as a bug in this page is welcome.

This site has no analytics either

These pages are static files. They run no analytics, set no cookies, embed nothing from another domain, and load no third-party fonts or scripts; the typefaces come from this site like everything else. A page can read an existing light or dark theme preference from your browser’s local storage under obadh-theme; that preference is not sent to the server.

The box you can type Bangla into runs the engine as WebAssembly inside your browser tab. It makes no request while you type and it sends nothing anywhere. What you type there stays in the tab and is gone when you close it.

The pages sit on someone else’s server, and that server keeps the ordinary request log any web server keeps: an IP address, a timestamp, the page requested, the browser string. Obadh adds nothing on top of that, reads none of it, and has no way to connect any of it to anything you typed.

Your writing stays here, even if suggestions learn from everyone

Autocorrect and next-word suggestions would get better if they could learn from how people actually write, and the only version of that worth building is one where nobody’s text leaves their device. Federated learning is the approach the project intends to take: models improve from updates computed on the device, and the writing itself stays where it was written.

None of it is built. There is no such code in any of the repositories today, nothing on your device is participating in anything, and there is no date. It is on this page because a plan that affects privacy should be visible before it exists, and because it comes with a condition: if it cannot be built so that no text, and nothing reconstructible into text, ever leaves the device, it does not ship. This page will describe it before the code is released.

Where to ask, and where to correct this page

Questions about anything on this page, and corrections to it, belong in an issue on the repository they concern: obadh_engine, obadh-ios, or obadh-macos. Issues are public, which is the point: a privacy answer given in private helps one person. The FAQ answers the shorter versions of these questions, and the contribute page says what else the project needs.

Last updated