⚡ DeepSeek Bubble

A floating AI assistant for Android, running entirely in Termux. A tiny draggable bubble sits over every app; tap it for clipboard quick‑actions (translate, summarize, explain, fix, reply) and an agentic chat where DeepSeek can run shell commands on your device to actually get things done.

Requirements

Three apps — install all from the GitHub releases (not the Play Store builds, which are outdated and incompatible):

• Termux
• Termux:GUI — the overlay engine
• Termux:API — clipboard, toast & text‑to‑speech

Termux packages:

pkg install termux-gui-bash jq curl termux-api

• termux-gui-bash — provides tgui-bash, which runs the script (required)
• jq, curl — API calls and JSON (required)
• termux-api — bridges to the Termux:API app for clipboard, toast, and speech

You also need a DeepSeek API key from https://platform.deepseek.com.

Setup

1. Clone the repo:

   git clone https://github.com/ExTV/termux-bubble
   cd termux-bubble

2. Grant Termux:GUI the "Display over other apps" permission (the script opens this settings screen automatically if it's missing).

3. Run it:

   chmod +x deepseek-bubble.sh
   ./deepseek-bubble.sh

4. On first launch, paste your API key and tap Save key. It's stored privately at ~/.config/deepseek-bubble/config (or set DEEPSEEK_API_KEY in the environment to skip this).

Usage

• Drag the bubble anywhere; tap it to open the panel.
• Clipboard actions — copy text, then tap Trans / Sum / Expl / Fix / Reply.
• Chat — type a request and Send. The model can run bash commands on your phone (read files, fetch things, run programs) to complete the task.
• While it's working, Send turns into Stop — tap it to cancel the current run. Collapse, drag, Copy/Speak and quit keep working during a run.
• Copy / Speak the last answer, Reset the conversation, ▽ collapse, ✕ quit.
• Mic dictates into the input box (Termux:API speech recognition).
• M switches the model between deepseek-flash and deepseek-v4-pro.
• Answers stream in as they are generated, and the console auto-scrolls to the newest line.
• The conversation is saved to ~/.config/deepseek-bubble/history.json and restored on the next launch; Reset clears it.

Configuration

Everything lives in ~/.config/deepseek-bubble/config (created on first save, chmod 600):

Key	Default	Meaning
DEEPSEEK_API_KEY	—	your key (an env var of the same name wins)
DEEPSEEK_MODEL	deepseek-flash	model id; deepseek-v4-pro is the bigger one
DEEPSEEK_API_URL	https://api.deepseek.com/chat/completions	any OpenAI-compatible endpoint
DEEPSEEK_MAX_TOKENS	8000	output cap per reply
DEEPSEEK_THINKING	off	off (fast, no reasoning pass), on, or default (server decides; for deepseek-flash that currently means on, so only off changes anything)
BASH_CONFIRM	0	1 = risky commands wait for a Run tap

Only DEEPSEEK_API_KEY is required. Every command the model runs is appended to ~/.config/deepseek-bubble/cmd.log with a timestamp, whether or not BASH_CONFIRM is on.

Because the endpoint is configurable, the bubble also works with a local or LAN model server (llama.cpp, vLLM, Ollama, OpenRouter, …) as long as it speaks the OpenAI chat-completions API and supports tool calling.

License

GPL-3.0 — see LICENSE.