BetterAIChat

简体中文 · English

A native Android AI assistant that uses your own API keys. Chat with leading LLMs, and let the AI operate your phone — open apps, take screenshots, search the web, run shell commands, and follow reusable Skills.

Unlike mainstream AI apps, BetterAIChat is a local-first agent: your API keys are encrypted on-device, there is no cloud, and the AI can actually do things on your device through function calling.

---

Screenshots

Conversations	Chat	Settings

---

Features

Chat & models

• Multiple providers: OpenAI-compatible (OpenAI, DeepSeek, Moonshot, Qwen, Ollama, any gateway), Anthropic Claude, Google Gemini
• Streaming replies with Markdown rendering, thinking-process display (reasoning stream), blinking cursor
• Agent picker: per-conversation Agent selection; built-in model catalog + server model fetch (/v1/models)
• Context usage tracking in the header (117.3K (12%))
• Context compression: summarize long conversations to free the window
• AI auto-titles for conversations; prompt templates (translate / summarize / polish / write / explain code / brainstorm)

Agents — one-tap setup, one per conversation

An Agent bundles a full configuration: provider + API key + model + temperature / max tokens / deep reasoning + optional custom system prompt.

• One-tap onboarding (4 steps): paste your API key → provider auto-detected from the prefix (sk-ant- = Claude, AIza = Gemini), with quick presets for DeepSeek / Qwen / Kimi / GLM / SiliconFlow → models fetched automatically → pick model → tune temperature / max tokens → default or custom system prompt
• Pick a different Agent per conversation (⋮ → Select Agent); manage all Agents in Settings → Agents
• API keys are encrypted with the Android Keystore; legacy settings auto-migrate to a default Agent

Modes (opencode-style)

Mode	Behavior
Chat	Pure conversation, no tools
Plan	Read-only analysis, write tools forbidden
Build	Default — every tool execution asks for confirmation
Max	Autonomous multi-step tool calls + deep reasoning (max effort / thinking)

The AI can operate your device

Tool	What it does
open_app	Launch any installed app
send_notification	Post notifications
set_brightness / set_volume / set_screen_timeout	Adjust system settings
set_flashlight	Torch control
take_screenshot	Screen capture
device_info	Device / battery / storage info
set_clipboard / get_clipboard	Clipboard read & write
set_alarm	One-shot reminders
schedule_repeat	Daily / weekly / hourly repeating reminders (manageable in Settings → Scheduled tasks)
speak_text	TTS read-aloud
web_search / web_read	Real-time multi-engine web search (Bing / Baidu / Brave / DuckDuckGo / Mojeek) and page reading
open_settings	Jump to system settings pages
run_shell	Root-level shell execution via Shizuku (pm, am, dumpsys, files…)
get_time	Current date / time / timezone
media_control	Play / pause / next / previous for active media sessions
set_ringer_mode	Switch ring / vibrate / silent mode
share_text	Share text or links via the system share sheet
open_dialer	Open the dialer pre-filled with a number
vibrate	Haptic vibration feedback
network_status	Connectivity, network type, WiFi name
get_foreground_app	Detect the currently active app
screen_ocr	Read on-screen text with per-line coordinates (screenshot + on-device Chinese/English OCR)
download_file	Download any file from a URL into the Downloads folder
ua_type / ua_tap / ua_swipe / ua_press	UI automation via Accessibility: type text, tap coordinates, swipe, press Home/Back/Recents/notification keys
ua_tap_text / ua_find_text	Find and tap UI elements by their text (more reliable than coordinates)
set_dnd	Toggle Do Not Disturb mode
manage_app	App management via Shizuku: force-stop / disable / enable / clear data / uninstall
write_document	Save AI-generated content as .md/.txt/.html to Downloads/Documents
set_wifi	Toggle WiFi (via Shizuku)
set_power_saver	Toggle battery saver (via Shizuku)
fetch_rss	Parse any RSS/Atom feed (titles, links, timestamps)
get_weather	Weather for any city (current + 3-day forecast, no API key)
calculator	Safe math evaluation: (12+5)*3, 2^8, %, decimals
generate_qr	Generate QR code images from text/links
keep_screen_on	Keep the screen awake for N seconds
create_automation	Create automations: time/battery triggers that run a sequence of tool actions unattended
list_automations / delete_automation	Manage automations
read_notifications	Read recent notifications (needs notification access)
get_screen_state	Screen on/off & lock state
list_installed_apps	App inventory with name/package/system flag, filter + limit
ping_network	Latency check for IP / domain / HTTP endpoints
get_exchange_rate	Live FX rates (no API key)
get_location	GPS / network location fix + accuracy
transcribe_audio	Record microphone audio and transcribe to text
ocr_file	OCR any image file (ML Kit, full build)
screen_record	Record the screen to MP4 (needs capture grant)
send_email	Compose an email via the system mail client
create_calendar_event	Create calendar events (opens the calendar pre-filled for confirmation)
get_app_usage	Per-app foreground usage stats (needs Usage access)
unit_convert	Unit conversion (length/weight/temperature/area/speed/data/time)
search_contacts	Search contacts by name (needs Contacts permission)
list_files / read_text_file	Browse / read files in Downloads & Documents

Screen analysis — the AI can see your screen

One tap → screenshot → vision model describes what's on your screen and gives operation advice. Works with any vision-capable model.

Automations — set-and-forget tasks

The AI can create automations that run tool sequences automatically:

• Time trigger: e.g. "every day at 22:00 silence the phone + enable DND"
• Battery trigger: e.g. "when battery drops below 20%, remind me to charge"
• Executed in the background via AlarmManager (no confirmation needed), completion is reported via notification
• Manage / toggle / delete them in Settings → Automations

Long-term memory

• The AI automatically distills key user information (name, preferences, agreements) into persistent local memory
• Memories are injected into every request, so the AI remembers you across sessions
• Auto-distills every 10 messages; manual trigger via ⋮ menu → "Distill important info (提炼重要信息)"
• Manage / delete memories in Settings → Memory (记忆)

Seamless context continuation

• When usage reaches 85% of the context window, the app warns and auto-compresses after the current turn
• Before compressing, the last 6 turns are saved as a snapshot — "Import recent chat" (导入最近对话) restores them as a message, so you continue without losing recent context

Full UI automation — the AI can drive your phone

With two toggles in Settings (Accessibility + Usage access), the AI gains a complete automation loop:

1. screen_ocr reads the text on screen
2. ua_tap / ua_swipe / ua_type / ua_press perform the actions (tap, scroll, type, Home/Back/Recents)
3. take_screenshot / screen_ocr verify the result

Works on any app — no root, no Shizuku needed for this path.

Skills (opencode-style)

• Import SKILL.md files with YAML frontmatter (name, description, allowed-tools)
• Skills can define their own tools with action types: alarm, notification, clipboard, intent, settings — including {param} templates
• Record actions into skills: after the AI completes a multi-step task, save the tool sequence as a reusable Skill
• Built-in action recorder, import/delete management

Themes

8 accent color schemes (橙/红/粉/靛蓝/蓝/紫/绿/青), default orange matching the app icon, instant switching, vibrant light mode.

Multi-engine web search

web_search merges Bing / Baidu / Brave / DuckDuckGo / Mojeek with URL dedup and 5-minute caching — reliable results even when one engine blocks scraping.

Attachments & documents

• Images (up to 4, auto-compressed, EXIF-corrected) sent to vision models
• PDF (on-device Chinese OCR), Word .docx, Excel .xlsx (all sheets), text files up to 1MB

Hands-free

• Voice assistant mode: AI speaks its reply, mic auto-opens, your spoken answer is sent automatically — a complete hands-free loop
• Voice input button, read-aloud for any message, auto read-aloud toggle

Organization

• Multi-conversation management: pin, archive, search (title + full-text content), rename, swipe-to-delete, clear context
• Starred messages with a dedicated favorites page
• Export conversations as Markdown (share sheet)
• Long-press message actions: copy / speak / edit & resend / star / delete
• Usage stats (conversations, messages, tokens, tool calls)
• Share-into-chat (ACTION_SEND) and deep link betteraichat://ask?text=…

---

Getting Started

Download

Download the APK from Releases — in a browser. The GitHub Android app corrupts large APK downloads; verify with the SHA-256 printed in each release, or use the smaller -lite build (~10 MB, no on-device OCR).

Configure

1. Settings → Agents → New Agent (or, right in a chat: ⋮ → Select Agent → New Agent)
2. Step 1: paste your API key — provider + Base URL are detected automatically (sk-ant- → Claude, AIza → Gemini), or pick a preset (DeepSeek / Qwen / Kimi / GLM / SiliconFlow) and tap Detect to fetch your server's model list
3. Step 2: pick a model · Step 3: temperature / max tokens / deep reasoning · Step 4: default or custom system prompt → save
4. Start a conversation with that Agent. In Build/Max modes, ask the AI to do things: "open the calculator", "remind me to drink water at 9am every day", "search today's news and summarize"…
5. Want another persona? Create more Agents and switch per conversation via ⋮ → Select Agent

Build from source

# JDK 17 + Android SDK (compileSdk 36)
./gradlew assembleFullDebug
# APK: app/build/outputs/apk/full/debug/app-full-debug.apk (lite variant: assembleLiteDebug)

---

Permissions

Permission	Purpose
Notifications	AI notifications & reminders
Microphone	Voice input & voice assistant mode
Camera	Flashlight control
Modify system settings	Brightness / screen timeout
Screen capture	Screen analysis (Android 15: choose "Entire screen" during consent)
Shizuku (optional)	Root-level shell execution

All permissions are granted manually; tools return clear errors when unauthorized.

---

Tech Stack

Kotlin · Jetpack Compose (Material 3) · OkHttp (SSE) · kotlinx.serialization · Room · Android Keystore · ML Kit (Chinese OCR)

:app         UI (conversations, chat, settings, stats, starred)
:core        models, engine, SSE, storage, DB, skills parsing
:providers   OpenAI-compatible / Anthropic / Gemini adapters
:skills      device tools + tool registry + action executor

Learning from this project

New to Android agent development? Read docs/HOW_IT_WORKS.md — a deep technical walkthrough covering the module architecture, the message pipeline, SSE streaming, the tool-calling agent loop, permission bridging, the automation engine, and UI rendering tricks (terminal-style scrolling, streaming cursor), with a suggested study order.

---

Roadmap

• Multi-model comparison (one question, several models side by side)
• Home screen widget
• Agent marketplace / shareable Agents

---

License

MIT