GPT Assistant for Android: the speech key ships in the repository
GPT assistant for Android, activated via volume keys for voice interaction, supporting features such as networking, taking photos, templates, parsing PDF and Office documents, and agent mode.
At a glance
- What is it?
- A Java Android app that is summoned with the volume keys, talks through four speech interfaces, browses the web through function calling, and can drive the phone through the accessibility service in an experimental agent mode. The default speech path uses an API key committed to the repository and protected only by a package-name check.
- Who is it for?
- GPT Assistant suits someone on a Huawei, Honor or Xiaomi handset who wants a system-wide assistant reachable without opening an app, and who is comfortable pointing it at a third-party relay to reach a model their network blocks. It does not suit someone who needs a hardware-independent compatibility claim, because the tested-device table is almost entirely Huawei-family and Xiaomi hardware plus one emulator.
- Can I use it commercially?
- Yes, with conditions. GPL-3.0 is a copyleft licence: if you distribute software that includes it, you must release that software's source code under the same licence. Running it internally without distributing it does not trigger that obligation.
- Is it still maintained?
- Activity is slowing. The repository last received commits 6 months ago.
- What is it written in?
- Mainly Java, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- Skythinker616/gpt-assistant-android909 ★ · +8
| Repository | 2026-09-10 | 2026-11-15 |
|---|---|---|
| Skythinker616/gpt-assistant-android | 901 | 909 |
8 Stars in the last 66 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
The default speech interface runs on the author's committed key
The app bundles someone else's credential and protects it with a package check. Four speech recognition interfaces are supported and the default is Huawei's, which the documentation describes as free, with the author's API key included in the app so it works out of the box, available for as long as the vendor's free period lasts. The key is protected, not hidden. The developer note explains that to prevent abuse the keys in the repository have package name and signature verification switched on, so anyone who builds from source finds Huawei speech recognition failing until they create their own AppGallery application and replace the authentication details, which means the `app/agconnect-services.json` file and the `hms_api_key` field in `app/src/main/res/values/strings.xml`. So the shipped build works because it carries the author's identity, and the source build deliberately does not.
Ten releases across two and a half years with two year-long gaps
The update log runs from 2023.09.10 to 2026.04.12 and lists ten entries: the first version with basic chat, Baidu voice input, TTS output and Markdown rendering; continuous conversation and GPT-4 support three days later; Huawei speech recognition in October 2023; networking that November; Vision image recognition in December 2023; advanced template syntax that same month; Google and Whisper speech recognition in January 2024; continuous voice conversation in May 2024; multiple document parsing in April 2025; and agent mode in April 2026. So the first eighteen months produced eight releases and the last two years produced two. The three visible tags are v3.0.0, v2.0.1 and v2.0.0, which means the entire 1.x maintenance line, visible in the tested-device table as versions 1.5.0 through 1.11.0, has no recent tag at all. The TODO list is three items: length limits on questions and replies, message compression, and downloading and installing updates from inside the app.
Networking is function calling plus a WebView with a fifteen second clock
Web access is not a search integration. The model issues a request through the function interface and the app fetches the page and returns the content, which has to be enabled in settings first. The fetch itself is an Android WebView, so the reachable set is whatever the phone's own browser can open. For most sites the model receives plain text only; for a named list of adapted sites it also receives search result links, and that list is Baidu, Bing, Google, Google Scholar, Zhihu, Weibo, JD, GitHub, Bilibili and CNKI. A page load times out at fifteen seconds, and a page that needs a login or a verification step simply will not be readable. The documentation also carries four warnings: the demonstrated results were produced with `gpt-3.5-turbo`; questions work better if you steer them with phrases like search or fetch online; failures are often the model choosing the wrong URL or the page changing; and because the page content is sent to the model, a larger model will burn a lot of tokens. Models without function support cannot use this at all.
The agent mode reads the accessibility tree, not the screen
The experimental mode is described as letting the AI read the controls on the current screen and perform taps, scrolls and text entry, which the documentation frames as enabling sending messages, ordering food and looking things up, without image recognition and with any model. What it actually reads is the accessibility tree, and the project's own troubleshooting entry says why some pages fail: some applications do not expose complete control information to the accessibility service, and some models do not support function or tool calling. So the limitation is not accuracy but visibility, which is a different failure shape. Two requirements are stated for turning it on, an accessibility service grant specific to agent mode and a model that supports tool calling. Two warnings follow, that recognition and operation are not guaranteed to be stable for every app or control, and that the app retains nothing of yours but payment, verification code and password screens are still advised against.
Volume key wake needs three grants and each has its own symptom
The headline interaction is a long press of the volume down key, and the troubleshooting section enumerates exactly what goes wrong when each prerequisite is missing. If a long press only changes the volume, the accessibility service is not enabled, and it may need re-enabling after a reboot. If the phone vibrates once and nothing appears, the background overlay permission is missing, and that permission may not exist on every device, in which case the note is to ignore it. If it works and then stops working after a while, the app is not allowed to run in the background. Three separate grants, three separate symptoms, and the documentation resolves each one without a diagnostic tool. There is a second entry point that avoids the accessibility service entirely, a button in the notification shade that opens the same interface and raises the keyboard. Conversation is continuous once you activate the chat icon, and individual messages can be deleted or retried from an avatar control.
The tested device list is Huawei, Honor, Xiaomi and one emulator
The compatibility table lists twelve combinations, and reading the manufacturers is more informative than reading the versions. Honor 7C on Android 8, Honor 20 on HarmonyOS 3 and 4, Mate 30 on HarmonyOS 3 and 4, Magic 4 on MagicOS 7, Magic 6 on MagicOS 8 and 9, Redmi K20 Pro on MIUI 12, Redmi K60 Pro on MIUI 14, Xiaomi 13 on MIUI 14, and a Pixel 2 emulator on Android 12. Every physical device in that table is from the Huawei family or Xiaomi, and the only entry from Google's own hardware is an emulator. The app versions tested range from 1.5.0 to 3.0.0, which is where the untagged 1.x line becomes visible. The repository reflects the same skew in a smaller way: an IDE project directory is committed at the root alongside the Gradle files, and the build is a plain Android Gradle project with a separate module for the speech recognition core.
Non-OpenAI model formats are explicitly declined
The application talks to whatever endpoint you configure in settings, filling in a base URL and a key, and the documentation covers two cases. The official case is the OpenAI endpoint with a key from an OpenAI account. The second case is a third-party relay, and the reason given is that the official service cannot be reached directly from inside China. A specific free relay project is recommended, offering a couple of model names with a limit stated as no more than two hundred requests per day per IP and key. What the project will not do is adapt each provider natively. The entry on non-OpenAI formats names Claude, Gemini and Ollama and says plainly that there is no plan to support them one by one, and points instead at two self-hosted gateway projects that convert several providers into the OpenAI shape. In other words, adding a model family is the user's deployment problem rather than an app update.
The privacy statement assumes the official endpoint
The privacy note is short and absolute: the program collects no personal information in any way, voice input goes directly to whichever speech interface you chose, questions go directly to the OpenAI API, and nothing passes through other intermediate servers. Each clause is true for the configuration it describes, and the third is the one to read twice. It holds when you point the app at the official endpoint. The recommended configuration for a network that cannot reach the official endpoint is a third-party relay, and in that path your questions do go through someone else's server, at a rate limit of two hundred requests a day. The first clause is also worth separating from the second: nothing is collected by the app, but the audio is transmitted to a speech vendor, and by default that transmission is authenticated with a key belonging to the author rather than to you.
Editorial conclusion
GPT Assistant suits someone on a Huawei, Honor or Xiaomi handset who wants a system-wide assistant reachable without opening an app, and who is comfortable pointing it at a third-party relay to reach a model their network blocks. It does not suit someone who needs a hardware-independent compatibility claim, because the tested-device table is almost entirely Huawei-family and Xiaomi hardware plus one emulator. Check four things first. Check which speech interface you will actually use, because the default one is the author's own key and its availability depends on a vendor's free tier lasting. Check the permissions the volume-key wake needs, which are three separate grants and each has its own failure mode. Check what you point it at, because the recommended configuration for networks that block the official endpoint routes your questions through a third party and the privacy statement assumes you do not. And check whether you want the agent mode, which reads the accessibility tree and can act on payment and password screens. The licence is GPL-3.0, the default branch was pushed on 2026-05-05, and the newest release is v3.0.0.
Frequently asked questions
Can ChatGPT control my phone?
Through this app's experimental agent mode, once you grant an accessibility service and use a model that supports function or tool calling. It reads the controls exposed on the current screen and can tap, scroll and type. Some apps hide their controls from the accessibility service, which is why some pages do not work.
How to use ChatGPT as an assistant?
Install the APK from the latest release, then fill in a base URL and API key in settings. Long press the volume down key to open the interface, hold it to speak, release and tap again to send, and replies are read back through the system text to speech engine.
How is GPT Assistant for Android activated?
Three ways. A long press of the volume down key, which needs an accessibility service grant, a background overlay permission and background running permission, each with its own symptom if missing. A button in the notification shade, which opens the same interface without the accessibility grant. Or by sharing or opening a file with the app from another application.
Which speech recognition interfaces does GPT Assistant support?
Four. Huawei is the default and ships with the author's key, Google requires its app installed and set as the system engine, Baidu needs its own app credentials and bills beyond a free quota, and Whisper works if your configured endpoint supports it. Speech output uses the system text to speech engine.
Does GPT Assistant for Android collect my data?
The privacy note says no personal information is collected, voice input goes directly to the chosen speech interface, and questions go directly to the OpenAI API without other intermediate servers. That last clause assumes the official endpoint rather than a third-party relay.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/skythinker616-gpt-assistant-android)Star history chart for your README
[](https://hysenlabs.com/projects/skythinker616-gpt-assistant-android)Was this useful?