Find your next open-source project

Search by name, purpose, or keyword — or paste a GitHub repository URL

Verlintas/NovaBAIC

AI Engineering

Local-first Android AI agent that calls 58 device tools to read the screen, automate UI, manage files and run scheduled tasks.

120
7
Kotlin
GPL-3.0
50 days old

What it is

BetterAIChat2 (codename Nova, formerly BetterAIChat) is a local-first AI agent app for Android. Instead of only chatting, the model runs through function calling and operates the device: taking screenshots, OCR, accessibility-based taps and typing, file reads/writes, web reading, reminders and scheduled jobs. It targets Android 8.0 (API 26) and above and has no GMS dependency.

With 58 built-in tools, the agent covers screen reading and control (ui_control can locate, act and assert in one call), system settings, content and network access, personal assistants, sub-agents, MCP remote tools and YAML skills. A bionic memory layer keeps identity aliases, timelines, provenance and versioned notes, and supports "do not remember" holds.

Four stock modes define how much autonomy the model gets: Chat (talk only), Chat+ (read-only tools, 8 rounds / 25 tool calls), Act (confirm each tool call), and Max (autonomous run with a 60-round / 160-call / 60-minute budget). Users can create, edit or restore modes.

Getting started

  1. Download and sideload the APK from GitHub Releases. The official site also bundles download channels, including an always-latest mirror.
  2. Open Settings → AI services, paste an API key (the provider is auto-detected) or select one manually, pull the model list and pick a model.
  3. Create a conversation and choose a mode above the input box. Start with Act so each tool call asks for confirmation before granting accessibility, screen capture or notification access under Settings → Permissions.
  4. To test device control, try "what is on my screen?" (screenshot + OCR) or "open WeChat and message File Transfer" (accessibility automation).

Optionally install Shizuku for run_shell and deeper app management, or pair wireless debugging once for background virtual-display capture without Shizuku.

Build from source

Building needs JDK 17 and Android SDK platform 37 with build-tools 37.0.0:

./gradlew :app:assembleDebug     # APK: app/build/outputs/apk/debug/app-debug.apk
./gradlew test                   # unit tests

For local development without an API key, two mock servers ship under dev/ (an OpenAI-compatible streaming server and an MCP server); forward their ports with adb reverse as described in the README.

Constraints and privacy

  • Tools only work after the corresponding Android permissions are granted; without them the agent returns a pointer to the right settings page rather than failing silently.
  • Screen capture is authorized per session and then reused. Background virtual-display keep-alive adds a persistent low-priority foreground service and raises battery use; it can be turned off.
  • web_research cannot handle password handoff or file upload/download, and screenshots of password, payment or verification pages are refused by default. Platform support varies by site.
  • API keys are encrypted with Android Keystore (AES-GCM) and stored locally; there is no cloud, telemetry or account. Plain HTTP is allowed only for loopback and private LAN addresses.
  • The README states the code is mainly written by an AI coding assistant, with requirements, review and release handled by the human maintainer; it invites independent review of correctness and safety.
  • Licensed GPL-3.0-or-later. Bundled offline place-name data is from GeoNames under CC BY 4.0.

HotRepo's take

No HotRepo ranking or star-growth snapshots are available for this repository, so no popularity trend can be inferred from our records. On the technical side, the project stands out for breadth rather than novelty of the chat UI: 58 tools, a documented Agent loop, mode-level budgets, skills and MCP in a single app. The main adoption barrier is also visible in the README: most capabilities depend on accessibility, screen capture and optional Shizuku or wireless-debugging pairing, so the useful surface for a new user depends heavily on how many permissions they are willing to grant. Because distribution is a sideloaded APK, readers who care about supply chain should compare releases against the open source tree and build it themselves.

Written by AI from the project's README, updated . The project's own docs are the source of truth.

You might also like

  1. wangzhishou/OneBoxAndroid AI agent and toolbox that runs 90+ in-app tools from natural language and opens 100+ file formats.Kotlin1,191
  2. guzhou079-arch/deepseek-harness-androidAn Android APK that bundles DeepSeek Harness with screen automation, virtual display, scheduled tasks and 27 skills, without Termux or root.JavaScript110
  3. wangzan101/MikaAn Android AI agent that operates apps and the web on a background screen to complete tasks you describe in natural language.Kotlin8,854
  4. totec448-spec/chat-on-steroidsA local MCP workspace that lets ChatGPT read files, run commands and delegate tasks across multiple agents.TypeScript5,182
  5. vercel-labs/agent-desktopA CLI and C library that lets AI agents read and operate desktop apps via the OS accessibility tree.Rust21,905