Giga Pisar Lands on Google Play: A Single Manifest Line Nearly Got It Banned

If you are building an Android app with AI inside, this story will save you some nerves. The author of "Giga Pisar", a voice-input app for Android, has finally published it on Google Play, and the app is now available in 178 countries. But a day before release, it nearly got branded as malicious. The cause turned out to be laughably small.

How one manifest line turned into a block

In GitHub issues, users asked for a mode in which Pisar works not with a cloud neural network but with a home server, for example Ollama. The idea is clear: the "brain" runs on your own hardware, and no data leaves. The server sits in the next room and usually has no certificate, so you have to talk to it over plain http.

To allow this, the author added a cleartext traffic permission to the manifest, that is, permission to transmit data without encryption. He built a version and put it on GitHub, and it would not install for people. Play Protect, Android's built-in antivirus, deemed the app malicious and blocked installation, even when connected by cable from a computer.

As the author writes, Google treats unencrypted traffic as a sign of a dangerous app and blocks such builds. He removed the permission, temporarily hid the home-server tab, and the new version installed without any questions.

What's new in Pisar itself

The main update is called "Shorten". Previously, the cloud "Brain" only polished the dictation: it removed "uh-huh" fillers and repetitions and added punctuation. Now you can speak the way people actually speak, at length and with repetitions, and the app will leave a short digest that is easy to read on a phone. The author noticed that long voice messages turned into text are hard to read: it is still a transcript of speech, not a written message. According to him, the mode is tuned so that the meaning does not change.

The prompts are open: in the settings you can see exactly what goes to the neural network, and you can rewrite the instruction to suit yourself. This is a nice touch for those who value transparency.

There are also small additions requested by users:

• sending a message with the volume-up key;
• working in Samsung Notes;
• the screen does not turn off during recording;
• the volume button works in Android Auto;
• voice cancel: "Pisar, cancel";
• a "No period at the end" toggle.

Filler sounds like "uh" and "um" are cut out right on the phone, without internet. Dictations of up to twelve words are inserted immediately and are not sent to the neural network, which saves both time and requests.

What this means for developers

The lesson is simple: if you plan to publish on Google Play, do not leave cleartext in the manifest, even for a harmless task like talking to a local server. Test the build on a device with Play Protect enabled before release, not on launch day. For local models this is an inconvenient limitation: the most private scenario turns out to be the riskiest for publication.

You will have to look for possible workarounds yourself. For example, set up TLS with your own certificate on the home server, or use a VPN tunnel. The author does not describe such solutions in the text, so these are only general ideas, not his plan.

There are caveats too. The story is told by a single developer, and it contains no official details on Play Protect's criteria: all we know is that after removing the permission, he solved the problem. So treat this as a practical observation, not a documented rule.

The project's plans are as follows: dictation history, a local "Brain" on the phone itself, and other languages. The order depends on what users ask for most often, and some tasks run up against smartphone capabilities. Feedback is collected in GitHub issues.