Network policy

The core typing engine has no network code anywhere in it. Prediction, autocorrect, gesture typing and layouts all behave the same with the radio off. A set of optional tools do reach the internet, each one only while you’re actively using it. Automatic backup runs on a timer too, if you set one up. This page is the complete list of what goes where, and Network activity shows every one of those requests as it happens, on your own device.

In the F-Droid edition every address on this page is a setting. WM Keyboard › Advanced › Servers lists each service and each download repository, and the same fields sit on each tool’s own page. Point one at a mirror or at a server you run and the keyboard calls that instead; leave a field empty and it calls the service’s own server. See Servers.

The Tools screen: every tool has its own Enabled switch, on by default.

What each tool sends, and when

Every network-touching tool makes a plain HttpURLConnection request. No analytics SDK, no ad network and no background sync sits underneath any of it.

Tool	Endpoint	Sends	Fires when
Translate	translate.googleapis.com (no key) or translation.googleapis.com (with a key)	the text you typed into the panel, up to 2,500 characters, plus the target language	as you type, about 400ms after you stop, only while the panel is open
Translate and DeepL Write, with your DeepL key	api-free.deepl.com (Free keys), api.deepl.com (other keys), or the DeepL server you entered	Translate: the same as above. DeepL Write: the text in the field or the selection, up to 2,500 characters. Both send your key in a header	only after you add a DeepL key under Tools / Translate. Translate as you type in its panel; DeepL Write only when you tap it
Grammar check, with LanguageTool on	api.languagetool.org (the public server), api.languagetoolplus.com with a Premium account, or the LanguageTool server you entered	the text in the field, cut into pieces of at most 18 KB on the public server, the language code, and your username and key if you set them. Never a password field or text typed in incognito	only after you switch LanguageTool on under Tools / Grammar check / Checkers (it starts off), and then while the Grammar panel is open, after each pause in typing (at least 1.2 seconds), for each check it is the one to answer: with Harper above it, that is every language but English. Fix all and the Fix grammar selection action send their text too. The public server is sent at most 20 requests and 75 KB a minute
Translate, with your own translation server	the server address you entered	the same as Translate above, inside a chat request with an instruction to translate, plus your key in a header if you set one	only after you enter a server address under Tools / Translate. Then as you type in the panel, in place of every other translation service
Translate, on-device models	dl.google.com (through Android’s download manager), not in the F-Droid build	nothing you typed: the request names the language model	when you press Download for a language, in the panel or under Tools / Translate
Translate, Firefox Translations models	firefox-settings-attachments.cdn.mozilla.net, Mozilla’s server for the files Firefox downloads, 64-bit phones with Android 9 and later	nothing you typed: the request names the model file	when you press Download for a direction, in the panel or under Tools / Translate
Translate, NLLB-200 model	huggingface.co, one fixed revision of a CTranslate2 conversion, 64-bit phones with 6 GB of memory or more	nothing you typed: the request names the model file	when you accept its licence and press Download, in the panel or under Tools / Translate
Add niqqud, Phonikud model	huggingface.co, one fixed revision	nothing you typed: the request names the model file	when you press Download under Languages → Hebrew → Hebrew options
Sticker generator, prompt translation	your online translators, the same as Translate above	the prompt, only when it isn’t English and no on-device translator could handle it	when you press Generate
Sticker generator, Improve with AI	your AI provider, as for the AI tools	the prompt, with an instruction to rewrite it	only when you press Improve with AI
Sticker generator, Pollinations	image.pollinations.ai (no key) or gen.pollinations.ai (with your key, in a header)	the prompt in English with the style’s words, in the address, plus size, seed and safety setting; then nothing more, the picture is the answer	only after you switch Pollinations on in the Engines list (it starts off), and then when you press Generate with it picked
Sticker generator, AI Horde	aihorde.net, then the picture from the Horde’s storage (*.r2.cloudflarestorage.com)	the prompt and negative prompt, size, seed, count and safety setting, and your Horde key or the anonymous one; then a status check every 4 seconds while it waits in the queue	only after you switch AI Horde on (it starts off), when you press Generate with it picked. Leaving the screen withdraws a queued request
Sticker generator, keyed services	api.openai.com, generativelanguage.googleapis.com, api.x.ai, api.stability.ai, api.bfl.ai, external.api.recraft.ai, api.replicate.com, fal.run, api.together.ai, router.huggingface.co, api.cloudflare.com, whichever you added; FLUX, Replicate and some others hand back a link to their own storage, which is fetched	the prompt in English, the negative prompt where the service reads one, size, seed, count, safety setting, the model name, and your key in a header	only after you add the service under Engines with your key, when you press Generate with it picked (or reached in One at a time). Test on its screen reads the model list or the account instead, and draws nothing
Sticker generator, your own server	the Automatic1111, Forge, SD.Next, ComfyUI or OpenAI-compatible address you entered	the same as the keyed services, plus a checkpoint name, and your key or user:password if you set one	only after you add the server under Engines, when you press Generate with it picked
GIF search	api.klipy.com, api.giphy.com, and gifcities.archive.org with web.archive.org once GifCities is on	your search text, or nothing for the trending feed	immediately when you open the panel (trending), then live as you type
Sticker search	same clients as GIF search	same	same
Reporting a GIF or sticker result	none of the app’s own	nothing: it builds a mail draft naming the provider, your search term, and the result’s id and URL	only when you tap Report in the long-press menu, and only your mail app sends it
Web search	api.search.brave.com	your search query	only when you submit a search, never as you type
Image search	api.search.brave.com	your search query	same (submit only)
Camera, search by photo	lens.google.com, www.bing.com, yandex.com or tineye.com, whichever Tools / Camera / Search site names, or your own server for Custom	the photo, scaled to at most 1280 px	only when you press Search on a photo with Search with set to Web. The default, Google Lens, and the share sheet hand the photo to another app and send nothing from the keyboard
Dictionary	kaikki.org, en.wiktionary.org, api.dictionaryapi.dev, api.urbandictionary.com if you turn it on, in the order set under Tools / Dictionary / Dictionary sources	one English word: your selection, or the word at the cursor	on panel open, if “Look up the word at the cursor” is on (the default); also on manual search. A source is asked only when the ones before it had no entry or could not be reached. Installed vocabulary packs are read on the device first, and send nothing; so are the offline dictionaries, wherever they sit in the list
Recorded pronunciations	en.wiktionary.org for the list of recordings, upload.wikimedia.org for the audio, and apifree.forvo.com only if you added a Forvo key	the word, and its language for Forvo	only when you tap an entry’s speaker in the Dictionary tool or the Look up card, with Recorded pronunciations on (the default). A recording already kept on the phone is played without asking
Synonyms	api.datamuse.com, kaikki.org, api.dictionaryapi.dev, in the order set under Typing / Suggestions / Synonym sources	one English word: the word you held on the suggestion strip	only when you tap Synonyms in the word menu, and a source is asked only when the ones before it had nothing. Installed vocabulary packs are read on the phone first, and send nothing
Wikipedia	wikipedia.org	your search text or an article title	only when you search or open an article, never on panel open
Smart home: Home Assistant	your own Home Assistant server, at the address you typed, over a WebSocket to /api/websocket and its web API; wmkeyboard.pages.dev/ha/ is fetched by your server once per browser sign-in, not by the phone	the key Home Assistant issued to the keyboard, in a header; the commands you give (a tap, a slider, an Assist sentence); a camera still when you ask for one. Back come the states of your entities	only while the tool is on and the home is signed in, and then while the panel is open or the keyboard is shown (your choice), plus a minute after. A snippet variable or the AI chat’s home tools connect for the moment they need
Smart home: Google Home	Google’s Home APIs through Play services on the phone, which reach Google’s service and, at home, the devices directly	what Google’s own consent screen granted: the devices of one home, read and commanded through Google’s SDK; nothing is sent by the keyboard’s own code	same as above; the SDK keeps its own connection to Google while the tool runs
Smart home: SmartThings	api.smartthings.com (Samsung’s cloud); wmkeyboard.pages.dev/st/ is where the browser lands once per sign-in, and it only forwards the result to the app	the key SmartThings issued, in a header; the device list with its states every ten seconds while the panel is open; the commands you give	same as above
Smart home: Hue bridge	the Hue bridge on your own network, over HTTPS with its own certificate pinned at pairing	the key the bridge minted, in a header; the commands you give. Back come the lights, rooms, scenes and sensor readings over the bridge’s event stream	same as above
Smart home: openHAB	your own openHAB server, at the address you typed	your API token, in a header; the commands you give. Back come the items and their state changes over the server’s event stream	same as above
Smart home: Smart Life and Tuya	the Tuya data centre you picked (openapi.tuyaeu.com, openapi.tuyaus.com and the like)	your project’s Access ID and an access token in headers, each request signed with the Access Secret, which never leaves the phone; the device list with its states every 30 seconds while the panel is open; the commands you give	same as above
Smart home: SwitchBot	api.switch-bot.com (SwitchBot’s cloud)	your token in a header, each request signed with the secret key, which never leaves the phone; each device’s state once a minute while the panel is open; the commands you give	same as above
Sync with another profile	no server, and not the network: 127.0.0.1 on the phone itself, on a port taken from the pairing code	to the copy of WM Keyboard in your private space or work profile only, after both prove they hold the same pairing code: this copy’s sync file, encrypted with a key made from that code	only while sync goes through an Other profile location, and only when both copies are running. In between it keeps one socket open on 127.0.0.1 and sends nothing
KDE Connect	no server at all: computers on your local network, over UDP 1716 and TLS on TCP 1716 to 1764	a name, a random device id and the list of features, to find computers; then, to a computer you paired, whatever you use the tool for: pointer moves, what you type there, the clipboard, files, media commands, battery level	never while nothing is paired and no device list is open. After that only while the link is allowed to exist (keyboard showing, by default), and to unpaired computers only while a device list is open. A paired WM Keyboard computer also gets the sync file, when Sync with paired devices is on. Album art for a player on the computer may be fetched from the web address that player names
Currency converter	open.er-api.com, falling back to api.frankfurter.dev. The F-Droid edition starts from cdn.jsdelivr.net (currency-api, which uses latest.currency-api.pages.dev when jsDelivr fails) and falls back to api.frankfurter.dev. The source is yours to pick from those, allratestoday.com and api.coinbase.com	nothing identifying: it fetches the whole rate table, converted locally	on panel open, or the first time a smart chip sees an amount. In the F-Droid edition that chip waits for your tap unless Fetch rates as you type is on. Cached for hours, not fetched per conversion
Internet speed test	speed.cloudflare.com by default; a LibreSpeed server you chose instead, your own or one from the volunteers’ list at librespeed.org, which is fetched when you open the picker	random bytes, down and then up, as many as the line carries in the time set; ten empty requests for the ping; one request that asks the server what network you are on, which the card shows. Nothing you typed	only when you press Start, never on panel open. A run moves tens to hundreds of megabytes, and Data saver asks first on a metered connection
Cryptocurrency prices	api.coinbase.com, falling back to cdn.jsdelivr.net; the F-Droid edition uses cdn.jsdelivr.net alone. api.coingecko.com is the other choice	nothing identifying: it fetches prices for the coins you have turned on, converted locally	only when you type a coin amount or select a coin on the panel, never on a plain panel open. In the F-Droid edition a typed amount waits for a tap on its chip. Cached for a few minutes
Weather	api.open-meteo.com, geocoding-api.open-meteo.com	the location you typed or searched, or nothing beyond a saved lat/long	on panel open, cached for a while, or when a weather chip needs a forecast (in the F-Droid edition that chip waits for your tap unless Fetch the forecast as you type is on); geocoding only fires while you’re editing the location
Location	nominatim.openstreetmap.org, or the Nominatim server you set in the F-Droid edition	the shared point at the precision you chose, and the keyboard’s language for the answer; a city-level share asks only for a city. Never the raw reading when a coarser precision is set	once per reading and precision, when the panel has a position: after Locate me, or on panel open if Android already allows location and Locate when the tool opens is on. Also once for Use my location on the weather and compass tools, to name the city. Turn Look up the address off and the tool sends nothing at all. Map tiles go only to a tile server you name yourself, none by default
AI (cloud provider)	the provider you picked, on its own API: Anthropic, OpenAI, Gemini, Grok (xAI), DeepSeek, or Brave (api.search.brave.com, which runs a web search with your text before it answers)	the text you ran the action on, plus a system prompt	only when you tap an action, never in the background
AI (any other OpenAI-compatible service)	whatever address you type in, for instance OpenRouter, Groq, Together or Mistral	same	same
AI (self-hosted Ollama or LM Studio)	a server address you enter on your own network	same	same, over plain HTTP
AI (on-device) Not on F-Droid	none	nothing leaves the phone	never: this provider has no network path at all
AI suggestions in the strip	your AI provider, as for the actions above	the last 120 characters before the cursor (40 to 500, as you set it), or only the last word; plus a short system prompt	off by default. When on, each time you pause about 0.6 seconds after a word ends; a key press cancels it. Never in a password field, a field that turns suggestions off, incognito, or while Data saver holds cloud AI back
Reporting an AI result	none of the app’s own	nothing: it builds a mail draft and hands it to your mail app	only when you tap Report, and only your mail app sends it
Clipboard link previews	whatever URL you copied	a GET request to that page	off by default; when on, right after you copy a new link
QR scan link previews Not on F-Droid	whatever URL a scanned code contains	a GET request to that page	off by default; when on, right after a QR code decodes to a URL
Voice typing (system recognizer)	Android’s platform speech service (not WM Keyboard’s own code)	your speech audio	whenever the on-device model for that language isn’t installed: always, below Android 12
Voice typing (Whisper)	none	nothing leaves the phone during dictation	never: fully offline once the model is downloaded
Voice typing (Whistle) Not on F-Droid	none	nothing leaves the phone during dictation	never: fully offline once the model is downloaded
Voice typing (Asr4All)	none	nothing leaves the phone during dictation	never: fully offline once a model is downloaded
Voice typing (Parakeet, Moonshine, Sherpa, Omnilingual, Vosk)	none	nothing leaves the phone during dictation	never: fully offline once the model is downloaded
Glide typing (FUTO Swipe engine)	none	nothing leaves the phone while you swipe	never: fully offline once the two models are downloaded
Voice typing (transcription server)	a server address you enter, on your own network or a service such as Groq or OpenAI	the recording of each phrase (up to 30 seconds, compressed as Opus at about 20 kbit/s, or 16 kHz WAV where the server cannot read that), the model name you set, the layout’s language unless you turn that off, and up to 150 of your own words plus your prompt text when words to listen for are on	once per phrase, after you stop speaking, only while you dictate with this engine
Music ID, online recognisers	api.audd.io, the ACRCloud host of your project, audiotag.info, shazam.p.rapidapi.com, or the recognition server you entered; each only once you switched it on with your key	the clip the microphone recorded when you pressed Listen (6 to 12 seconds, Opus at 20 kbit/s or 16 kHz WAV; the Shazam listing gets the last five seconds raw) and your key for that service. Never in a password field or incognito; held back by data saving’s Song recognition online row	once per Listen, to each recogniser in turn until one names the song; not at all when the song was read from the player on the phone
Music ID, cover art	itunes.apple.com, then api.deezer.com	the title and artist of a song already recognised, to fetch its cover and a store link; never the clip	once per song that came without a cover, and only when data saving allows online recognition outright

The AI panel’s Report button is the one row that quotes your own text without any network of its own. It opens a draft in your mail app holding the generation and the text it ran on, and you read and send it, or don’t. See Reporting a result.

Some rows need a note. GIF and sticker search fetch a trending feed the instant you open the panel, before you’ve typed anything. See GIFs for that behavior in full. Currency, weather and location are the tools that touch the network without you typing a query first. None sends anything you typed. Currency pulls a public rate table and converts locally, weather sends only the location you already saved or searched, and location asks for the address of the point it is about to share, at that share’s precision. Voice typing’s network path isn’t literally WM Keyboard’s own HTTP code, since it’s Android’s SpeechRecognizer used as a silent fallback, but audio still leaves the device that way. See Voice typing for the full breakdown, or Offline voice (Whisper) for the fully on-device alternative.

Photo backgrounds are not a toolbox tool, but they do reach the network, so they belong in this list. Open the picker, type a word, and that word goes to Unsplash and Pexels along with the subject, color and orientation filters. Nothing else about you goes with it. Thumbnails load from each service’s own servers rather than being copied through anything of ours, which is how a photographer is credited with the view. Choosing a photo downloads that one file at keyboard size, and tells Unsplash once that it was taken, which is how its download counts stay honest. A rotating background is the one part that fetches without you watching. It tops up a small set of photos as the keyboard closes, never while you type, and skips that entirely while the device is locked, under power saving, on mobile data unless you allow it, and when High contrast keys is on. See Online photos and rotation.

Everything else in the toolbox is offline by design: OCR (apart from the one-time download of a language’s Tesseract or PaddleOCR data, below, and the Online model engine, which sends each scanned photo to the OpenAI-compatible service you name and is off unless you choose it), QR/barcode scanning, document scanning, the calculator, unit converter, symbols picker, password generator, QR code generation, grammar and spell-check, the camera tool (apart from searching with a photo on the Web setting, above), and every calendar/alt-calendar/qibla/moon-phase/solar-time feature. See Scanner and Utility tools for their own confirmations of that.

The services behind them

AI tool settings: a self-hosted Ollama or LM Studio server address, with the plain-HTTP disclosure below it.

“Google Translate” is two different endpoints. With no key, WM Keyboard calls translate.googleapis.com, the same unofficial public endpoint browser translate extensions use. That’s fine for personal volumes, but it’s rate-limited and carries no uptime guarantee. Add a Cloud Translation API key and it switches to translation.googleapis.com, the official, billed Google Cloud API. See Translate for that switch in detail.

The full set of third-party hosts a WM Keyboard install can reach, beyond what’s already in the table above:

• api.github.com, on a build downloaded from the project’s GitHub releases, is where the update check asks what the newest release is. It runs at most once every six hours, and only while you have the settings app open. The request carries nothing but a version string identifying the app, and it is a conditional request, so an unchanged list of releases sends no data back and costs nothing against the limit GitHub applies to everyone sharing your IP address. Press Download and the APK comes from github.com (which redirects to objects.githubusercontent.com), and raw.githubusercontent.com serves the release notes when you press What’s new, reading one markdown file out of the tag you are being offered and nothing else. That request is made only when you press the button. Play Store and F-Droid builds never make any of these requests. The notes for the version already installed ship inside the app. The one exception, on any build, is a version released before its notes were written: opening them then reads that version’s file from its tag on raw.githubusercontent.com. See Updating the app.
• f-droid.org, on an F-Droid build and no other, is asked what version F-Droid has, on the same six-hour schedule. That build downloads nothing: it can only point you back at the F-Droid client.
• api.keyman.com and downloads.keyman.com are asked for a Keyman layout’s typing rules, only when you press its Typing rules row or accept the offer to fetch them as you switch the layout on. The request names the keyboard and nothing else: the first asks for its current version, the second serves the package.
• raw.githubusercontent.com/wasi-master/wmkeyboard-data is the project’s own data repository, source for downloadable dictionaries, word-pair (n-gram) packs, emoji keyword packs, CJK conversion tables, the 5 MB background remover the sticker editor fetches on a phone that has no Google Play services to supply one, and a copy of Keyman keyboards’ typing rules that is read only when keyman.com can’t serve them. See Downloads below.
• raw.githubusercontent.com/wasi-master/wmkeyboard-addon-repository is the sample addon repository, and also where the one-tap Noto Color Emoji font on the Emoji screen is fetched from (about 10 MB, plus a few hundred bytes of version manifest beside it). That one tap doesn’t require adding a repository first.
• raw.githubusercontent.com/wasi-master/wmkeyboard-monkeytype-sounds is the second pre-added addon repository, a set of key-sound packs. Like the one above it is removable, and removing it is permanent.
• raw.githubusercontent.com/wasi-master/wmkeyboard-community-addons is the third, themes and layouts other people sent in. Removable the same way. Sending one of your own opens a GitHub issue in your browser; the app itself sends nothing. See Sharing with everyone.
• raw.githubusercontent.com/tesseract-ocr/tessdata_fast Not on F-Droid is the Tesseract project’s own language data, pinned to its 4.1.0 release. The text scanner fetches one file per language, 0.4 MB to 12 MB, only when you press Download for it. The request names the language and nothing else.
• raw.githubusercontent.com/iFleey/PPOCRv5-Android holds PaddleOCR’s PP-OCRv5 models converted for LiteRT, pinned to one commit. The text scanner fetches its three files, 11 MB together, only when you press Download for Chinese or Japanese where ML Kit can’t read them. The F-Droid build fetches only the list of characters from here, and the two models, in MNN’s format, from the data repository’s models release. The requests carry nothing about you.
• huggingface.co is the source for offline Whisper, Whistle, Asr4All, Parakeet, Moonshine, Sherpa and Omnilingual speech models, the FUTO Swipe glide models, the NLLB-200 translation model, the Phonikud niqqud model and on-device AI models. Most are public. A couple of gated on-device models need your own Hugging Face account token.
• alphacephei.com, Alpha Cephei’s site, holds the small Vosk models. Offline Vosk fetches one zip, about 40 MB for most languages, only when you press Download for that language. The request names the file and nothing about you.
• qaihub-public-assets.s3.us-west-2.amazonaws.com, Qualcomm AI Hub’s public bucket, holds Stable Diffusion 1.5 compiled for each Snapdragon NPU. The sticker generator fetches the one zip for your chip, about 710 MB, only when you pick that model and accept its licence, on a Snapdragon 8 Gen 3, 8 Elite or 8 Elite Gen 5. The request names the file and nothing about you.
• fonts.gstatic.com serves Google’s Noto animated emoji. Long-press an emoji that has an animated version and the keyboard fetches a small WebP to loop in the popup, then fetches the full GIF when you tap Send. Nothing about you goes with either request. It only happens when Animated emoji is on and you’re in a field that accepts images. See Animated emoji.
• Google’s downloadable-font provider (com.google.android.gms.fonts) supplies the Noto face the keyboard draws every non-Latin script in, and the Noto Color Emoji font choice arrives the same way. Play services makes that request, not the app, and nothing about you travels with it: a font is asked for by family name only. It happens the first time a script needs a face the device hasn’t already cached, without you choosing anything. A device with no Play services falls back to the system font per glyph instead.
• Your backup locations, if you add any. That’s Google Drive (www.googleapis.com), Dropbox (www.dropbox.com to sign in, then api.dropbox.com, api.dropboxapi.com and content.dropboxapi.com), OneDrive (login.microsoftonline.com to sign in, then graph.microsoft.com), Box (account.box.com to sign in, through wmkeyboard.pages.dev/oauth/ on the way back, then api.box.com and upload.box.com), pCloud (api.pcloud.com or eapi.pcloud.com, and the download host it names), MEGA (g.api.mega.co.nz and the storage hosts it names), iCloud Drive (idmsa.apple.com to sign in, setup.icloud.com, and the iCloud Drive hosts that names), Proton Drive (drive-api.proton.me and the block storage hosts it names), or the S3, WebDAV, SFTP, SMB, NFS, FTP, Git, IMAP, Synology or QNAP server whose address you typed in yourself. A Nextcloud WebDAV location can also sign in through your own server’s login page. Nothing is contacted until you add a location, so none of these are contacted on a fresh install. What travels is your backup file, and your sync file if you turn on sync between devices, and nothing else. See Backup & restore.
• Whatever addon repository you add. Three are pre-added the first time you open the Addons screen, the three named above, but adding a reference doesn’t fetch anything by itself. See Installing addons for when a repo actually gets contacted.
• Ollama and LM Studio are whatever address you point them at, with no fixed host. Traffic to a self-hosted server is plain HTTP, and the app allows cleartext traffic app-wide to support it, so keep that server on a network you trust. See AI tools for the full setup.

None of these services receive anything from WM Keyboard beyond the single request described for that tool, and there’s no shared analytics identifier tying them together. Automatic backup is the one thing here that fires without you, and only to the destination you chose yourself.

Keeping API keys safe

WM Keyboard › Tools › Translate

Each tool that needs a key has its own field, all blank by default: Translate’s Cloud Translation and DeepL keys, GIF search’s KLIPY and GIPHY keys, Web and Image search’s Brave and Tavily keys, photo backgrounds’ Unsplash and Pexels keys, and the AI tool’s per-provider keys (Anthropic, OpenAI, Gemini, Grok, DeepSeek, Brave, and whatever key the OpenAI-compatible service you pointed it at wants) plus its Hugging Face token for gated on-device models, and Voice typing’s transcription server key. Leave any of them blank and that tool either falls back to a free endpoint (Translate) or shows a “needs an API key” state (GIF, search, AI). See each tool’s own page under Tools for exactly where its field lives.

Every key you enter lives in the same place as the rest of your settings: a plain, unencrypted Android DataStore preferences file, protected by Android’s normal app-sandboxed storage rather than by any extra encryption WM Keyboard adds on top.

Keys are excluded from a manual backup export unless you turn on Include API keys, and from Android’s own device backup path they’re excluded unconditionally. See Secrets stay out by default for the full export mechanics, including the exact warning text the app shows you.

One gap

Ollama and LM Studio server addresses aren’t treated as secrets the way API keys are. They’re not excluded from a manual export, even with “Include API keys” off. That’s usually a LAN address, but if you’ve folded a token or password into that URL yourself, it travels in plain exports. Keep that in mind before sharing a backup file.

Downloads: dictionaries, models, and addons

Every downloadable asset comes over HTTPS, but the integrity check behind it isn’t uniform. Don’t assume every download is checksummed just because some are.

A large file doesn’t come down as one long request. Models, feature libraries, conversion tables, spellers, interface language packs and app updates are fetched as up to twelve requests at once, each asking for its own piece of the file. GitHub, which hosts most of them, holds a single connection from Bangladesh to between 0.3 and 0.8 MB a second, and twelve side by side measured about 5 MB a second on a phone; a server that is fast on one connection, such as Maven Central, is no slower for it. Each request is its own row in Network activity, to the same host as before. The pieces are written into a .part file with a small .ranges file beside it that records how far each piece got, so a download that stops, whether you cancelled it, lost the signal or the app was closed, asks next time only for the bytes it doesn’t have. A server that won’t answer a request for part of a file is read in one go instead. None of this changes the checks in the table: a file is still measured and hashed whole before it is used. Word lists, word-pair data, emoji keywords and the other small packs that are unpacked as they arrive still come as one request, because a size tier stops reading part way through the file and fetching all of it first would cost more data.

Download	Source	Integrity check	Confirms before using mobile data?
Dictionaries and wordlists	the project’s GitHub data repository	none beyond the gzip stream’s own format check	Only from the “download everything for this language” button, which asks while Data saver is on and its Downloads row is set to Ask each time (the default)
Emoji keyword packs	same data repository	none beyond decode success	Same button, same setting
Offline dictionaries for the Dictionary tool	same data repository	the file must carry the header of the dictionary it was asked for before it replaces anything	Only while Data saver’s Downloads row asks
CJK conversion tables (pinyin, kana, stroke, cangjie, jyutping, and the Traditional phrase table), the RIME engine’s schema packs, the Mozc engine’s dictionary and zinnia’s handwriting models	same data repository	SHA-256 verified before the file is used	No
Keyman typing rules	downloads.keyman.com, or the same data repository’s copy when keyman.com can’t serve them	the file must parse as a keyboard before it’s kept, and the data repository’s copy is also SHA-256 verified against the checksum listed beside it	No
Keyman layouts (GitHub and Google Play editions), about 2 MB	the data repository’s layouts release on GitHub. Fetched by itself once a Keyman layout is on, and tried again when the phone comes back online	SHA-256 of the contents, checked against the value built into the app before the file is used. The server is not trusted at all	Only while Data saver’s Downloads row blocks it on this connection
Offline Whisper speech models	Hugging Face	SHA-256 for whisper.cpp models, from a fixed revision; a file-size floor only for LiteRT models	Yes, around 150 MB. See Offline voice (Whisper)
Offline Whistle speech model Not on F-Droid	Hugging Face, pinned to one revision	SHA-256 verified before the file is used	Only while Data saver’s Downloads row asks. See Offline voice (Whistle)
Offline Asr4All speech models	Hugging Face, pinned to one revision per size	SHA-256 verified for both files before the model is used	Only while Data saver’s Downloads row asks. See Offline voice (Asr4All)
FUTO Swipe glide models	Hugging Face, futo-org/futo-swipe pinned to one revision	SHA-256 verified for both files before they are used	Only while Data saver’s Downloads row asks. See The FUTO Swipe engine
Offline Parakeet and Moonshine speech models	Hugging Face, pinned to one revision each	SHA-256 verified for every file before the model is used	Only while Data saver’s Downloads row asks, and Parakeet’s 719 MB also asks on mobile data. See Parakeet and Moonshine
Offline Sherpa speech models, Silero VAD and the English punctuation model	Hugging Face, pinned to one revision each	SHA-256 verified for every file before the model is used	Only while Data saver’s Downloads row asks, and a model of 150 MB or more also asks on mobile data. See Offline voice (Sherpa)
Phrase end detectors: Silero VAD for Asr4All, Smart Turn v3.2, and FireRedVAD	Hugging Face for Silero and Smart Turn, pinned to one revision each; FireRedVAD from FireRedTeam’s own GitHub repository, raw.githubusercontent.com, pinned to one commit. Fetched from the rows under Phrase end detector, never on their own	SHA-256 and exact size of every file before it is used	Only while Data saver’s Downloads row asks
Offline Omnilingual speech model (Meta’s Omnilingual ASR 300M) and Silero VAD	Hugging Face, pinned to one revision	SHA-256 verified for every file before the model is used	Only while Data saver’s Downloads row asks, and its 367 MB also asks on mobile data. See Offline voice (Omnilingual)
Offline Vosk speech models Not on F-Droid	alphacephei.com, one zip per language, each a fixed file name	SHA-256 of the zip, built into the app, checked before it is unpacked; no entry may land outside the model’s folder	Only while Data saver’s Downloads row asks, and a model of 150 MB or more also asks on mobile data. See Offline voice (Vosk)
Text scanner languages (Tesseract) Not on F-Droid	the Tesseract project’s tessdata_fast 4.1.0 on GitHub	none beyond the file arriving at the length the server announced; a file Tesseract can’t load is not read	Yes, from the scanner’s button and from the settings page. See Scanner
NLLB-200 translation model	Hugging Face, JustFrederik/nllb-200-distilled-600M-ct2-int8 pinned to one revision	SHA-256 and exact size of each of its four files before the model is used	Only while Data saver’s Downloads row asks; its licence is asked for first. See Translate
Firefox Translations models (on-device translation)	Mozilla’s attachment server, firefox-settings-attachments.cdn.mozilla.net, the files Firefox’s own Remote Settings point at	SHA-256 and exact size of each download, and again of each file after it is unpacked, against values built into the app	Only while Data saver’s Downloads row asks. See Translate
Handwriting models for other scripts (PaddleOCR), where ML Kit has no model for the language	the data repository’s models release on GitHub, 7.8 to 13.4 MB per script. Fetched from the handwriting panel’s button or the settings row, never on its own	SHA-256 and exact size of both files, checked against values built into the app before they are used	No
Text scanner models for Chinese and Japanese (PaddleOCR)	PPOCRv5-Android on GitHub, pinned to one commit; on the F-Droid build, where they read Latin-script text too, the two models (21 MB) come from the data repository’s models release instead	SHA-256 of each of the three files	Yes, from the scanner’s button and from the settings page. See Scanner
On-device AI models	Hugging Face, per model	no checksum, a file-size floor only	Yes, at 500 MB. See AI tools
Search by meaning model (EmbeddingGemma 2)	the data repository’s models release on GitHub, about 155 MB for text and 220 MB more for pictures. Fetched when you pick it, never on its own	SHA-256 and exact size of every file, checked against values built into the app before it is used	Only while Data saver’s Downloads row asks
Sticker generator models: SDXS and the picture filter	the data repository’s models release on GitHub, about 970 MB together. Fetched when you pick the model and accept its licence	SHA-256 and exact size of every file, checked against values built into the app before it is used	Only while Data saver’s Downloads row asks
Sticker generator NPU model (Stable Diffusion 1.5)	Qualcomm AI Hub’s bucket, qaihub-public-assets.s3.us-west-2.amazonaws.com, the zip for your chip from release 0.64.0	SHA-256 and exact size of the zip before it is unpacked	Only while Data saver’s Downloads row asks
Upscaler networks (waifu2x UpConv7, one for drawings and one for photos)	the data repository’s models release on GitHub, 2.2 MB each. Fetched the first time you use Upscale in the sticker editor or the photo editor, never on its own; a small sticker sent to WhatsApp or Messenger uses the drawing network only once it is already here	SHA-256 and exact size of each file, checked against values built into the app before it is used	No
Neural suggestions model (IndicXlit)	the data repository’s models release on GitHub, 13 MB for every language at once. Fetched when you turn the setting on for a language, never on its own	SHA-256 and exact size of every file, checked against values built into the app before it is used	Only while Data saver’s Downloads row asks
Romanized dictation model (IndicXlit, Indic to Latin letters)	the data repository’s models release on GitHub, 13 MB for every language at once. Fetched when you tap its row on a language’s page, never on its own	SHA-256 and exact size of every file, checked against values built into the app before it is used	Only while Data saver’s Downloads row asks
Addons (themes, icon packs, plugins, sounds, layouts, sticker packs)	whichever repository you added	SHA-256 verified when the manifest supplies one, and mandatory for plugins specifically, since they run code	No
Signal sticker packs	Signal’s sticker storage, cdn-ca.signal.org. No Signal account is involved, and the request names the pack and nothing about you. The signalstickers.org gallery that the same screen links to opens in your browser, and the keyboard itself never contacts it	every file is checked against the key in the pack link (HMAC-SHA256) before it is opened, so a file that was changed on the way is refused	Yes. The pack preview asks before its first request while Data saver is on and its Downloads row is set to Ask each time (the default)
Telegram sticker sets	Telegram’s Bot API, api.telegram.org, through the bot you made. Every request carries your bot token, which is how the Bot API knows which bot is asking, and the name of the set. Your own Telegram account is never involved	none beyond the files arriving under the size cap; each is opened as an image, a Lottie animation or a video and dropped if it is none of those	Yes. The set preview asks before its first request while Data saver is on and its Downloads row is set to Ask each time (the default)
Rboard themes	the Rboard repository on GitHub, raw.githubusercontent.com/GboardThemes/PackRepoBeta. The list is fetched when you open Browse Rboard, and a pack when you press it. The Google Play build has no Browse Rboard and never makes this request	each pack is SHA-256 verified against the checksum in the list before it is opened; the list itself is capped at 2 MB	Yes. The screen asks before its first request while Data saver is on and its Downloads row is set to Ask each time (the default)
Feature libraries (GitHub and F-Droid editions)	the native code behind offline voice typing, the compression of recordings sent to a transcription server, grammar and spell checking, the RIME engine for Chinese, the Mozc engine for Japanese, the scanners (text, QR and the built-in document scanner), handwriting (ML Kit’s, zinnia’s for Chinese and Japanese, and MNN for the other scripts without ML Kit), on-device translation and on-device AI, fetched the first time each feature is used, or all at once from About / Feature libraries. The project’s own libraries come from the data repository’s natives release on GitHub. Google’s ML Kit, LiteRT and LiteRT-LM libraries come from Google’s Maven repository, dl.google.com, and Qualcomm’s NPU runtime, Microsoft’s ONNX Runtime (behind Offline Sherpa and Neural suggestions) and Alpha Cephei’s Vosk library (behind Offline Vosk, GitHub edition only) from Maven Central, repo1.maven.org. The keyboard reads only the one file it needs out of each package. The F-Droid edition downloads only after you turn on Download feature libraries, and the NPU runtime only after you accept Qualcomm’s licence. The Google Play edition gets the same parts from Play instead and never makes these requests, and the offline build makes none at all: it takes the same files as imports	SHA-256 and exact size, checked against values built into the app before the file is used. The server is not trusted at all	Only while Data saver’s Downloads row asks
Addon repository manifests themselves	the repository URL you added	none: it’s the trust root, so there’s nothing to check it against. Capped at a fixed size instead	No
App updates (GitHub builds only)	this project’s GitHub releases	SHA-256 verified against the checksum published in the release, and then the APK’s signing key is compared against the key your install is signed with. A release with no checksum is refused rather than installed on trust. The small .dm baseline profile fetched alongside it is checked against the same checksum file and skipped if it does not match	Yes, and it names the size before it starts

Only CJK packs, feature libraries, Keyman layouts, the search by meaning and neural suggestions models, app updates, the data repository’s copy of Keyman rules, Rboard theme packs, and addons with a declared hash get cryptographic verification. Wordlists, emoji packs, Whisper models, and on-device AI models rely on HTTPS transport integrity plus a basic size check rather than a checksum, and Keyman rules from keyman.com on HTTPS plus the check that they parse. That matters if you’re deciding how much to trust a third-party addon repository versus the project’s own data feed. See Installing addons for how that mandatory plugin checksum is enforced.

Google components, and what they report

Everything above is WM Keyboard’s own code making its own request. Three Google libraries the app is built on top of talk to Google on their own schedule, and this is the list of them. None of them can see what you type. None of them can be switched off from inside the app either, so the choice is whether to use the feature.

Component	In which build	Reaches Google when	What it sends
ML Kit (handwriting, OCR, QR/barcode scanning, document scanning, sticker background removal, on-device translation)	every build but F-Droid’s	when you use one of those features	Per Google’s own disclosure: device make, model, OS build and available ML accelerators; the app’s package name and version; a per-install identifier; latency figures, the API settings and input/output sizes; the SDK version, the kind of event (init, model download, detection) and any error codes. Google describes this as diagnostics and usage analytics, says it’s encrypted in transit and not passed to third parties, and provides no switch to disable it.
Downloadable fonts	every build with Play services present	the first time a script’s default keyboard font is needed, and whenever you pick a Google font in the theme editor	the font family name and this app’s package name, asked of the Play services font provider. Play services makes the actual request, not this app.
Play Store libraries (in-app updates, and the on-demand downloads of the on-device AI runtime, the on-device translation engine, the interpreter behind offline voice typing and the sticker editor’s own background remover, the handwriting recogniser, the libraries behind the whisper.cpp, Asr4All, Parakeet, Moonshine and Whistle voice engines, ONNX Runtime behind the Sherpa voice engine, the compressor for transcription server recordings, the scanner’s offline text reader, and the grammar checker)	Play Store builds only	on the update check, when you first use an on-device AI model, when you first pick on-device translation, when you first use offline voice typing or a background removal without Play services, when you first download a handwriting model, when you first dictate with one of those voice engines or a transcription server, when you first download a text-reader language pack, and when you first open the grammar tool or the spell checker starts	the app’s package and version, handled by the Play Store app that installed it

On a Play Store build the update check is the only thing that library is asked, and it is asked only when the app can see that the Play Store installed this copy. Sideload a Play-flagged APK and the app doesn’t ask at all, because the answer can only ever be a refusal.

The same builds carry Play’s review library, under the same condition. It is asked for the rating card at most once every 120 days. Your word count stays on the phone: the app reads it to decide whether to ask, and sends only the request. Any rating or review you leave goes to Google Play from Play’s own card, and the app never sees it.

Google Drive as a backup destination adds a fourth, but only if you connect it: signing in is a Google sign-in, and after that the app talks to Drive directly over ordinary HTTPS. Nothing is connected by default.

The F-Droid build ships no ML Kit at all, so the first row doesn’t exist there, and the F-Droid build carries no Play Store libraries either. The font row is the one that survives everywhere. It’s an AndroidX API talking to whatever font provider the device has, so a phone with Play services (or microG) on it has that path no matter which build you installed. A phone without one falls back to the bundled Noto faces and never asks anybody anything.

Options

WM Keyboard › Tools

• Per-tool Enabled switch (default on for every tool, including the networked ones). Turning a tool off here, or from the “Enabled” toggle at the top of that tool’s own settings screen, removes it from the toolbar, the toolbox, and its hardware leader-key letter all at once. A disabled tool can’t be opened, so it can’t reach the network.

WM Keyboard › Tools › Power saving

• Background network is on by default whenever Power Saving itself is active, and Power Saving is off by default. It stops the fetches that happen without you asking for them: clipboard link previews, QR scan link previews, and the dictionary’s look-up-on-selection. Every tool you open by hand keeps working normally, translate, search and AI included. See Power saving for the complete list of what else it drops.

WM Keyboard › Advanced › Data saver

• Data saver is the network-shaped version of the same idea, and the one to reach for if what you’re protecting is a data allowance rather than a battery. Each feature that fetches gets its own answer: allow it, ask first, or turn it off while the connection is metered. The fetches that happen on their own default to off, and the ones you start yourself default to asking, so a GIF grid on mobile data explains itself and offers a Use mobile data button rather than coming up empty. See Data saver for the full table.

There’s no single “airplane mode for the keyboard” switch beyond those. The closest equivalent is turning off every networked tool individually, or turning every Data saver row off and leaving the manual switch on. The typing engine itself needs no switch at all, because prediction, autocorrect, gesture typing and layouts have no network code to disable in the first place.

Details & edge cases

• You can build a keyboard with no internet permission at all. Build it with -Pwmkb.noInternet=true and the permission is left out of the APK. Everything on the device keeps working, and every tool on this page fails the way it does offline. See Building.
• The in-app Privacy screen lists the tools, not the fetches. Its note names translate, weather, web search, image search, GIFs, stickers, dictionary lookup and the AI tools, and says the network is used only at the moment you press one. That covers the deliberate half. It leaves out the things that fetch on their own (currency chips, language data, link previews) and the download screens, which is what this page adds. See Data collection for the same note read from the storage side.
• AI’s default provider needs a key before it does anything. Claude (Anthropic) is the pre-selected provider chip, but with a blank key, so the tool shows a “needs an API key” state out of the box rather than silently trying a provider that isn’t configured.
• Currency and weather cache their results rather than fetching on every open. Currency holds for a few hours, coin prices for a few minutes, and weather for about fifteen minutes, so most opens after the first don’t touch the network at all.
• Coin prices are fetched lazily. Cryptocurrency is on by default, but the coin source is only asked when you actually type a coin amount or pick a coin on the panel. If you never do, your keyboard never talks to it.
• Only one tool reads your field rather than its own search box. Translate, Wikipedia and both search tools send what you type into the panel, and nothing else: the translate panel is its own window and never looks at the field behind it. Dictionary is the exception, reading the selected or cursor word when auto-lookup is on.
• Addon repository manifests aren’t fetched proactively. Seeding the two pre-added references doesn’t contact them. A manifest only downloads when you actually open or refresh the Addons screen, which is also the visit that seeds them.
• Cleartext traffic is allowed app-wide, not scoped to Ollama and LM Studio specifically, which is what makes plain-HTTP requests to a self-hosted AI server possible at all. It doesn’t weaken HTTPS anywhere the app picks the address: every tool in the table above is contacted over HTTPS, and an addon repository URL on plain http is rejected rather than quietly upgraded. The places a plain address is your own choice are the self-hosted AI server, a Home Assistant home address typed without https:// (which is how most installs answer on the LAN), and an S3 backup endpoint or a Git backup server you type with http://.

See Data collection for what the keyboard learns from you locally, and Backup & restore for how API keys and the rest of your settings move between devices.