橘子岛 (Orange Island)

An independently maintained fork of Agora — a BYOK LLM client with multi-provider access, agentic workflows, and remote device control.

English | 中文

Download

• Build from Source — Clone the Orange Island repository and build with Android Studio (see Getting Started).

---

橘子岛 — a BYOK Android client for AI power users. Connect to built-in providers and custom endpoints with your own keys, branch conversations non-linearly, run models locally via llama.cpp, and use remote shell tools. Conversations are stored locally; configured providers and tools receive the data needed for the features you use.

Project Origin & Credits

橘子岛 (Orange Island) is an independently maintained, modified version of Agora, originally developed by newo-ether. The application has been renamed and further developed by 小橘、猫猫.

This fork is not an official Agora release and is not affiliated with or endorsed by the upstream project or its author. The original copyright notice and MIT license are preserved in LICENSE; see NOTICE for attribution and modification notes.

Why 橘子岛?

• Configurable connections: Model requests go to the provider or proxy you configure. Chat history is stored locally in a Room database; remote services apply their own data handling policies.
• Non-Linear Thought: A tree-structured message database lets you edit any past message, regenerate responses, and explore alternative branches without losing context.
• Agentic by Default: Multi-round tool calling with web search, code execution, remote file operations, memory management, and semantic conversation search.
• Remote Control: Manage servers, edit files, and search code through Conch. Configuring an API key enables its application-layer encryption; use HTTPS for transport security.

Features

Multi-Provider Access

• 8 built-in providers: OpenAI, Anthropic, Google Gemini, DeepSeek, Qwen (DashScope), OpenRouter, Ollama, Local (GGUF via llama.cpp)
• Unlimited custom providers with arbitrary base URLs and API keys
• BYOK: Bring your own API keys — no subscriptions, no middlemen
• Multiple API keys per provider with named aliases for easy rotation
• Per-provider base URL override for proxies and self-hosted endpoints

Agentic Tools

• Web Search — DuckDuckGo Lite (anonymous, no key), Brave, Serper, Tavily, and SearXNG integration
• Code Execution — Gemini code execution for running and testing code inline; Alpine Linux sandbox via PRoot with SAF file access
• Image Generation — BYOK text-to-image via OpenAI-compatible /v1/images/generations, rendered inline in chat
• Remote Shell & File I/O — Execute commands, read/write/edit/glob/grep files on remote servers via the Conch protocol
• Memory — Persistent active memory and saved memory files across conversations
• Conversation Search — RAG-powered semantic search over chat history

Thinking & Reasoning

• Deep reasoning: OpenAI o1/o3, Anthropic extended thinking, Gemini thinking, DeepSeek-R1, Qwen QwQ
• Configurable thinking level (low/medium/high)
• Streaming think-tag renderer with collapsible UI and duration tracking

On-Device Intelligence

• Local LLM inference via llama.cpp — run GGUF models entirely offline
• Local embeddings for on-device semantic search (RAG)
• Ollama provider for self-hosted models on your local network

Remote Device Control (Conch Protocol)

Conch application-layer encryption requires an API key. A blank-key endpoint sends plain JSON and should use HTTPS.

• ECDH key exchange + AES-256-GCM encryption + HMAC-SHA256 signing
• Token bucket rate limiting and nonce-based anti-replay protection
• Multi-device support — configure and switch between multiple remote servers
• MCP integration — Conch as a Claude Desktop MCP server

Knowledge Management

• RAG semantic search across all past conversations using cosine similarity
• Configurable similarity threshold and keyword/model search methods
• Selectable embedding model (remote or local), independent of chat model
• Context window management with real-time token counting and sliding window
• Visual context rollout indicator dims messages outside the active window

Data Portability

• .oi Export/Import: Conversations, memories, prompts, settings, and API keys in one portable file
• Merge, Replace, and Skip import strategies
• Auto Backup — periodic WorkManager-based backup with configurable period, categories, and retention
• Third-Party Import: Claude and ChatGPT export formats (.zip / .json)
• API key safety warnings for both export and import workflows

Customization

• System prompt templates with three-section editor (system prompt + user prepend + user append)
• Variable substitution: {sent_time}, {sent_date}, and extensible variable system
• Per-conversation model and system prompt switching
• Per-message model selection from the chat bottom bar
• Per-conversation generation overrides (temperature, max tokens, penalties)
• Auto title generation with configurable model

UI & UX

• Modern Material 3 design in Jetpack Compose with dynamic color (Material You)
• Light / Dark / System theme modes with configurable color schemes
• Non-linear branching: Edit any past message and branch into alternative conversation paths
• Real-time streaming with message anchoring and animated auto-scrolling
• Haptic feedback throughout the UI (long-press, selection, success/error)
• Immersive gesture-driven image and media viewer
• Markdown rendering with syntax highlighting, LaTeX math, and code blocks
• Image, video, PDF, and file attachment support with thumbnails
• iOS-style collapsing large-title in settings with shared page transition animations
• Blur effects with configurable performance toggle
• English, Chinese, and Traditional Chinese language support

Documentation

📖 User Manual — In-repository documentation covering installation, providers, tools, search, memory, shell, and more.

Getting Started

Prerequisites

• Android Studio (Ladybug or newer recommended)
• Android SDK 34+
• A valid API key from a supported provider

Quick Setup

① Launch Open 橘子岛 on your device.

② Settings Open Settings from the nav bar.

③ API Key Select a Provider and add your API Key.

④ Models Models → "Sync from All Providers."

⑤ Customize System prompts, context, search, memory.

Running Local Models

① Place Put a GGUF model file on your device.

② Import Settings → Provider → Local → "Import GGUF Model".

③ Configure Set context size, temperature, and other parameters.

④ Select Choose your local model from the chat picker.

Setting Up Remote Shell (Conch)

① Deploy Deploy the Conch server on your target machine.

② Add Device Settings → Shell Devices → add URL and API key.

③ Use The model auto-discovers shell devices for commands, files, and search.

Tech Stack

• Language: Kotlin
• UI Framework: Jetpack Compose (Material 3, dynamic color)
• Architecture: MVVM with Kotlin Coroutines & Flow
• Local Storage: Room Database with tree-structured message schema & DataStore Preferences
• Networking: OkHttp with SSE streaming
• Serialization: kotlinx.serialization
• Native: llama.cpp via Android NDK (CMake) for on-device LLM inference and embeddings
• Image Loading: Coil
• Markdown: Multiplatform Markdown Renderer M3
• Math: JLaTeXMath-Android

Privacy

Conversations and settings are stored locally. Messages, selected attachments, and relevant context are sent to the model providers, proxies, or tools you use. Remote services may retain logs under their own policies. Update checks contact the configured release service; crash reports are submitted only after user confirmation. Exports and backups can contain conversations and credentials. See the Privacy Policy.

License

The application code is distributed under the MIT License, preserving the original copyright notice of newo-ether. Third-party code and assets retain their respective licenses; the MIT badge does not relicense those components.

Builds with the Linux sandbox include PRoot (GPL-2.0-or-later), talloc (LGPL-3.0-or-later), and Alpine packages under several licenses. See NOTICE, the in-app third-party license pages, and source distribution instructions for attribution and release requirements. Downloaded models and user-installed plugins have their own terms.

For Orange Island support, use this repository's Issue tracker or the in-app feedback channel. Please do not direct support requests for this fork to the upstream author.

Open Source Licenses

Component	License	Copyright	Source
Orange Island (this fork, all first-party code)	MIT	(c) 2026 Orange Island contributors	this repository
Agora (upstream project this fork is based on)	MIT	(c) 2026 newo-ether	NOTICE
llama.cpp — on-device LLM inference	MIT	(c) 2023-2026 The ggml authors	thirdparty/llama.cpp
PRoot (Termux fork) — Linux environment	GPL-2.0-or-later	(c) STMicroelectronics; patches by Termux	thirdparty/proot
talloc (Samba) — memory allocator	LGPL-3.0-or-later	(c) Andrew Tridgell, Stefan Metzmacher	thirdparty/talloc
JLaTeXMath — LaTeX rendering	GPL-2.0 with Classpath Exception	(c) Scilab Enterprises / opencollab; Android port by Dimitry Ivanov	bundled via ru.noties:jlatexmath-android
Gradle Wrapper	Apache-2.0	Gradle contributors	gradlew
All other dependencies (Kotlin/Compose/OkHttp/etc.)	Apache-2.0, MIT, BSD, EPL, ISC, etc.	their respective authors	listed in-app under Settings → About → Third-Party Open Source Licenses
Alpine sandbox packages	Per-component GPL, MIT, BSD, MPL, Apache, Zlib, etc.	Their respective authors	Source distribution