# Privacy policy (F-Droid edition) > The formal policy for the F-Droid build: no Google library compiled in, nothing collected, and every network call named. URL: https://wmkeyboard.pages.dev/privacy/policy-fdroid/ Section: Privacy & security **Effective 20 October 2026.** This policy covers the WM Keyboard Android app (`com.wasimaster.wmkeyboard`) as built and distributed by F-Droid, and this documentation site. WM Keyboard is written and published by Wasi Master, an individual developer. The F-Droid build and the others are not the same app underneath, so they get one policy each rather than one policy with a page of footnotes. This page is the F-Droid build, which F-Droid compiles from source. It is compiled with `wmkb.enableFdroid=true`, `wmkb.enablePlayStore=false` and `wmkb.enableGms=false`, which means no Google library is on its compile classpath at all. Up to 0.6.2 this was the `lite` edition, which GitHub published too. If you installed from Google Play or GitHub, you have a build that does carry Google libraries. Read [Privacy policy (Google Play edition)](https://wmkeyboard.pages.dev/privacy/policy/) instead. The app tells you which one you have under About: the version line says F-Droid on this one. ## The summary WM Keyboard collects nothing. There is no account, no advertising, no tracking, no analytics, no crash reporting, and no server belonging to this project that the app sends anything to. What you type is processed on your device and stays there. In this build there is no caveat to that. The other builds have to disclose Google's ML Kit reporting its own diagnostics; this build does not contain ML Kit, so there is nothing of that kind to disclose. The one Google-adjacent code path that survives is downloadable fonts, and it is inert on a device without Google Play services. See [What this build does not contain](#what-this-build-does-not-contain) below for the exact list. The app does not sell, rent or share personal information, because it does not receive any in the first place. ## What this build does not contain These are compiled out, not switched off. They are absent from the APK, and you can check that against the [F-Droid build recipe](https://github.com/wasi-master/wmkeyboard/blob/main/fdroid/com.wasimaster.wmkeyboard.yml) and the source it builds from. | Absent | What it would have done | What you lose | |---|---|---| | **Google ML Kit** | reports device and app information, a per-install identifier, and timing and error data to Google whenever one of its features runs | handwriting input, OCR, QR and barcode scanning, document scanning, sticker background removal | | **Google Play in-app updates** | asks the Play Store whether a newer version exists | nothing: F-Droid's own client handles updates | | **The self-updater** | downloads an APK from the project's GitHub releases and installs it | nothing, and it could not have worked here anyway: F-Droid signs its builds with its own key, so an APK from GitHub cannot install over an F-Droid install. This build has no `REQUEST_INSTALL_PACKAGES` permission and no installer code. What it has instead is a check that asks F-Droid, described below | | **Play Feature Delivery** | downloads a feature's native library as a separate module | nothing: this edition downloads those libraries itself, after you agree to it (see below) | | **Google Play services auth** | gets an OAuth token for Google Drive | Drive as a backup destination. Local storage, WebDAV, an S3-compatible bucket, an SFTP, SMB, NFS or FTP server, a Synology or QNAP NAS, a Git repository, a mail account, Dropbox, OneDrive, Box, pCloud, MEGA, iCloud Drive and Proton Drive all still work | | **LiteRT-LM** | runs a language model on the device | the LiteRT-LM models (`.litertlm` and `.task` files). GGUF models still run on the device through llama.cpp, which this edition builds from source, and cloud AI providers still work with your own key | | **Qualcomm QNN runtime** | runs an on-device AI model on a Snapdragon phone's NPU, under Qualcomm's own licence | the **NPU** choice under Compute, which this edition could not use anyway without LiteRT-LM | | **LiteRT Whisper models** | offline dictation on Google's LiteRT | the older LiteRT Whisper models. Offline Whisper on whisper.cpp and the other offline voice engines still work | The native libraries for offline voice typing (whisper.cpp, ExecuTorch and sherpa-onnx), for grammar and spell checking (Harper and Hunspell), for the RIME engine for Chinese (librime), the Mozc engine for Japanese (Mozc), on-device translation (Bergamot), neural phonetic suggestions (IndicXlit) and GGUF models for on-device AI (llama.cpp) are built from source with the rest of this edition, but they are not in the APK. The keyboard downloads one the first time its feature needs it, from the project's data repository on GitHub, and only after you turn on **Download feature libraries** under Privacy. F-Droid did not build or review those downloaded files, which is why the setting is off until you turn it on, and why it asks you to confirm. Before the keyboard uses a file, it checks the file's SHA-256 against a value built into the app. Until you turn the setting on, those features say that they need it. Only one Google code path is left. The keyboard can ask Android's downloadable-font provider for a font: some scripts' default keyboard face, and every "Google font" choice in the theme editor, come from there rather than from the APK. That provider is part of Google Play services. On a device that does not have Play services or microG, the request is never made and the keyboard falls back to the Noto faces bundled in the app. On a device that does have it, Google Play services makes the request and receives the font family name and this app's package name. Nothing about you goes with it, and the app never sees the network call. It asks the system for a font and gets one back. ## What stays on your device The keyboard writes several things to its own private app storage. None of it is transmitted anywhere by the app. - **Learned words and word pairs.** Words you use that were not already in a dictionary, plus the bigrams, trigrams and skip-grams (the word two or three back paired with the one that followed) used to predict and rank the next word. - **Clipboard history**, if you have the clipboard tool enabled, along with anything you pinned. - **Typing statistics**, such as counts of keystrokes and words, if you have statistics enabled. - **Your settings**, themes, layouts, snippets and personal dictionary entries. - **Downloaded data**, such as dictionaries and emoji keyword packs. You can delete all of it. Learned data is one confirmed tap on the app's Privacy screen, and clipboard history and typing statistics each have a clear action of their own on the storage screen under About. Uninstalling removes everything the app ever wrote. Password fields, and any field an app marks as sensitive, are excluded from learning, suggestions and clipboard capture automatically. Incognito mode applies the same exclusions to everything for as long as it is on. ## What leaves your device, and only when you ask Typing itself never uses the network. A set of optional tools do, each one only while you are actively using it. This is the same list as the other builds minus the rows that need code this build does not have. | Tool | What it sends | To | |---|---|---| | Translate | the text in the panel | Google Translate's public endpoint, or Google Cloud Translation with your own key | | Web and image search | your query | Brave Search, or Tavily if you add a Tavily key | | GIF and sticker search | your query | KLIPY, GIPHY | | Dictionary | one word | `kaikki.org`, `en.wiktionary.org`, `api.dictionaryapi.dev`, `api.urbandictionary.com` if you turn it on, in the order you set | | Synonyms, from the word menu | one word | `api.datamuse.com`, `kaikki.org`, `api.dictionaryapi.dev`, in the order you set | | Encyclopedia | your query | Wikipedia | | Weather | a location you entered | Open-Meteo | | Currency and crypto rates | nothing identifying | currency-api (`cdn.jsdelivr.net`) and Frankfurter, or ExchangeRate-API, Coinbase or CoinGecko if you pick one | | Photo backgrounds | your search terms | Unsplash, Pexels | | AI writing tools | the text you ran an action on | the provider you chose, using your own API key, or a server you run yourself | | AI chat's Deep research and Summarize | the search queries the model planned, and an ordinary request for each page read; the pages' text then goes to your AI provider with your question | your web search provider, and the websites of the pages read | | Grammar check, with LanguageTool switched on | the text in the field, its language, and your LanguageTool username and key if you set them. Never a password field or text typed in incognito. Nothing at all with Harper, the default | `api.languagetool.org` (LanguageTool's public server), `api.languagetoolplus.com` with a Premium account, or a LanguageTool server you run yourself | | Sticker generator, with an online engine switched on | your prompt in English with the style's words, the seed, the picture size and count, and your key for that service. Nothing at all with On this phone, the default | the engine you picked: `image.pollinations.ai` or `gen.pollinations.ai`, `aihorde.net`, `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`, or a Stable Diffusion, ComfyUI or OpenAI-compatible server you entered; the pictures are fetched back from that service's storage | | Voice typing | your speech audio | Android's own speech service | | Language and model downloads | nothing about you | GitHub, Hugging Face, and Mozilla for the Firefox Translations models | Under default settings, typing on its own reaches none of these. A weather or currency smart chip waits for your tap before it fetches anything, and language data downloads only when you accept the prompt or press the button. Each has its own switch. Every service in the table can also be pointed at another server, including one you run yourself, under Advanced › Servers. [Network policy](https://wmkeyboard.pages.dev/privacy/network/) documents every one of these in full, including what is cached and for how long, and [Permissions](https://wmkeyboard.pages.dev/privacy/permissions/) covers what each Android permission is for. Every one of them is an ordinary HTTPS request made by this app's own code. You can read that code: there is no SDK in between, and the whole tree is public. When you use one of these tools, your data is handled by that service under its own privacy policy, not this one. The app ships no key for the AI providers: those only work if you supply your own. The default services those tools talk to are not free software. Every one of them is opt-in, and the keyboard is fully usable with all of them untouched. Most also have a free, self-hostable alternative you can switch to above. Voice typing is the row to read carefully. Until you set up one of the offline engines (which needs **Download feature libraries** on and a model downloaded), dictation goes through Android's own `SpeechRecognizer`. On most devices that is a Google component and your audio may leave the device. If that matters to you, use an offline engine or do not use voice typing. ### The update check One request in this build is not a tool. About once every six hours, and only while you have the settings app open, it asks `f-droid.org` which version of WM Keyboard F-Droid has. If that is newer than yours, a card offers to open F-Droid. It downloads nothing and installs nothing; the F-Droid client does that, the same as it would without this check. Turn the card off with **Ask me about updates** under About, and the check stops asking on its own initiative. It asks F-Droid rather than GitHub deliberately, so the version you are offered is one you can actually install from where you installed this. ## Permissions This build asks for the same Android permissions as the others, and every one belongs to a single optional feature that is off until you turn it on: microphone for voice input, camera for the camera tool, contacts for name suggestions, calendar for the calendar tool, photos for the clipboard's screenshot capture, notification access for media controls and one-time-code chips, usage access for the clipboard's "show source app". Each shows a plain explanation inside the app before Android's own prompt appears, and refusing leaves the rest of the keyboard working. The full list is on [Permissions](https://wmkeyboard.pages.dev/privacy/permissions/). ## Backups Backups are optional and written to the destination you choose. That can be local storage, a cloud drive you connect, a WebDAV server, an S3-compatible bucket, an SFTP, SMB, NFS or FTP server you run, a Synology or QNAP NAS, a Git repository you own, a folder in your own mail account, or a Box, pCloud, MEGA, iCloud Drive or Proton Drive account. Google Drive is not offered in this edition. By default the file is plain JSON. Turn on a backup passphrase and it is encrypted with AES-GCM under a PBKDF2-HMAC-SHA512 key. Nothing is uploaded anywhere by default and no copy is sent to the developer. See [Backup and restore](https://wmkeyboard.pages.dev/privacy/backup/). ## Children WM Keyboard is not directed at children and collects no data from anyone, including children. There is nothing to delete on request because nothing is ever received. ## Third-party components The app bundles open-source libraries and data (dictionaries, emoji annotations, fonts). None of them phone home. The full list, with licences, is in the app under About, Licences and attributions, and the app is MIT licensed. ## Changes If this policy changes, the effective date at the top changes with it, and the history of every edit is public in the [repository](https://github.com/wasi-master/wmkeyboard/commits/main/docs/src/content/docs/privacy/policy-fdroid.mdx). ## Contact Questions about this policy: `arianmollik323@gmail.com`, or open an issue at [github.com/wasi-master/wmkeyboard/issues](https://github.com/wasi-master/wmkeyboard/issues).