Miffan

A native Android AI client that brings models, assistants, tools, and local or remote workspaces together.

English · 简体中文 · 繁體中文

Miffan is an open-source AI workspace designed for Android. Connect the model services you already use, give different assistants their own prompts, memories, tools, and personalities, and keep conversations and files organized in one native app.

Use an API key with OpenAI-compatible, Gemini, or Claude services, or sign in with an eligible ChatGPT subscription for Codex access. Miffan does not bundle a model or replace a provider account; availability and charges depend on the services you configure.

What makes Miffan different

• One home for different models. Mix official APIs, compatible gateways, self-hosted endpoints, and a Codex subscription without rebuilding your workflow around one provider.
• Assistants are real workspaces. Each assistant can have isolated prompts, model parameters, memory, tools, MCP servers, Skills, visual identity, and conversation history.
• The phone can do more than display chat. Miffan can search the web, work with files, run a local Linux workspace, connect to remote servers over SSH, use device capabilities, and expose the same conversations through a browser.
• A character system with purpose. Choose customizable Miffan bowl characters or the blue whale girl, with expressions that respond to conversation state and time of day.

Remote workspaces · new in 3.4

Connect your own server over SSH and let an assistant work in a chosen project directory. Keep local projects and remote machines in the same app, with clear identities and connection states.

• Multiple hosts and projects. Manage SSH hosts, credentials, and project directories, including machines on a Tailscale network your device has already joined.
• Files and an interactive terminal. Browse, preview, edit, import, and export remote files, or open a persistent terminal to run commands yourself. Returning from a file preview keeps your current directory.
• AI that works where your files live. Bind a workspace to an assistant and use file and shell tools there. AI shell permissions and execution confirmations are managed separately for local and remote targets.
• SSH key management. Generate or import keys, copy public keys, and view or export private keys, with optional passphrase encryption for backups.

Open Workspaces → New → Remote to choose a host and project directory, then bind the workspace in the assistant settings. The one-time introduction also links directly to workspace management.

Local and remote projects	SSH host management	Discover the new workspace

Screenshots use demo hosts and projects; the displayed UI language is Simplified Chinese.

Remote workspaces currently support Linux/Unix SSH/SFTP hosts. Miffan uses an existing network connection and does not set up Tailscale or keep tasks running autonomously while the phone is offline. See the remote workspace guide for setup and current limits.

Meet the blue whale girl

Miffan and the blue whale girl · character illustration

蓝色大肥鱼 is a cheerful, rice-loving companion alongside the original Miffan bowl characters. Open Settings → Appearance to preview the collection and try it with a dedicated assistant.

• An assistant of her own. The first trial creates a separate whale-girl assistant and opens a new chat. Later trials reuse it and preserve your edits. Her name, personality prompt, model, and tools remain editable; existing assistants and conversations stay intact.
• Expressions that follow the conversation. She smiles, enjoys petting, looks proud, reacts to updates, eats while waiting, chews during text output, sways with spinning eyes during actual reasoning, and dozes with a breathing nose bubble. Native drawing keeps the background transparent and supports reduced motion.
• A coordinated appearance. Vivid blue hair, fins, and a bow come with light and dark palettes, onboarding previews, and startup artwork. The bowl and whale launcher icons can be selected independently. Restoring the previous palette keeps the dedicated assistant.
• Easy to discover. New installations can choose the collection during onboarding; existing users receive a one-time introduction on the chat home screen. You can also choose the whale avatar for an individual assistant.

Screenshots

Living mascot	Character styles	Tool-aware chat

Selected-text translation

Select text in any Android app and choose Miffan-Translate from the text action menu (the label follows the app language). Miffan shows the translation in a compact floating window without taking you away from the current page.

1. Select text and choose Miffan-Translate	2. Review or copy the translation

Features

Models and providers

• OpenAI Chat Completions and Responses-compatible services, Google Gemini and Vertex AI, and Anthropic Claude-compatible services
• Browser-based OpenAI Codex sign-in using access included with eligible ChatGPT subscriptions
• Built-in presets for popular official services and gateways, plus fully custom providers, base URLs, models, request paths, headers, and body parameters
• Model discovery and configurable modality, reasoning, tool-use, context-window, and generation settings
• Authenticated HTTP/SOCKS5 proxy support, custom User-Agent, connection testing, and optional provider balance queries
• Chat, reasoning, tool calls, image generation, and multimodal input according to the selected model's capabilities

Conversation experience

• Streaming responses, message editing and regeneration, response branches, favorites, folders, and local full-text search
• Per-conversation system prompts, history compression, automatic titles, follow-up suggestions, token usage, and generation statistics
• Images and document attachments; local text extraction for PDF, DOCX, PPTX, and EPUB when needed
• Rich Markdown and HTML rendering with syntax highlighting, LaTeX, tables, Mermaid diagrams, images, and diffs
• Export conversations as Markdown or images, share content into Miffan, and hand shared text to a selected assistant

Assistants and Miffan characters

• Independent assistant configuration for models, prompts, sampling parameters, context limits, request customization, and chat backgrounds
• Memory, references to recent chats, preset messages, quick messages, regular-expression transforms, mode injections, and lorebooks
• Import Tavern character cards in JSON or PNG format
• Four Miffan character kinds, six curated palettes, three motion profiles, and optional app-theme color synchronization
• Semantic animations for idle, thinking, success, error, typing, submitting, attention, and time-of-day scenes, with reduced-motion support

Tools and extensions

• MCP over SSE or Streamable HTTP, including OAuth flows and per-assistant server selection
• Install and manage Skills from files, GitHub repositories, and the Skill.sh catalog with guarded install targets
• Web search through Bing, Tavily, Exa, SearXNG, Brave, Perplexity, Firecrawl, Jina, Grok, and other services, plus custom JavaScript search adapters
• Optional local tools for time, clipboard, JavaScript, text-to-speech, user questions, screen time, and calendar events
• Local Linux and remote SSH workspaces with file management, an editor, interactive terminals, working-directory context, and AI file/shell tools

Voice, translation, and browser access

• Configurable speech recognition through OpenAI Realtime, DashScope, Volcengine, MiMo, and Step
• Android system speech plus configurable TTS services including OpenAI, OpenRouter, Gemini, MiniMax, Qwen, Groq, xAI, MiMo, ElevenLabs, Fish Audio, and Step
• In-app AI translation and Android selected-text translation in a compact floating window
• Optional local web server for browser access on the device or LAN, with password authentication, localhost-only mode, and mDNS discovery

Data and portability

• Conversations and settings are stored in the app's local database
• Selective local backup and restore, backup reminders, WebDAV, and S3-compatible storage
• Import provider settings and conversations from Chatbox, provider settings from Cherry Studio, and compatible backups from RikkaHub
• Independent application ID, signing identity, release channel, and deep links; Miffan can coexist with RikkaHub on the same device

See the feature compatibility matrix for the complete provider, search, speech, tool, and portability coverage.

Download and first setup

Miffan currently distributes signed APKs through GitHub Releases. Official releases target arm64-v8a devices running Android 8.0 or newer.

1. Install the latest Miffan APK.
2. Open Settings → Providers and configure a service, or sign in to OpenAI Codex with a supported subscription.
3. Add or discover a model, then select it globally or for a specific assistant.
4. Enable search, MCP, Skills, local tools, or a workspace only when you need them.

Miffan uses the application ID me.ayuilos.miffan.app and the miffan:// deep-link scheme. Existing RikkaHub data is not shared automatically; export a backup from the old app and import it into Miffan if you want to migrate.

Beginning with 3.0.0-rc.1, Miffan uses an independent SemVer release line that does not encode a RikkaHub version. Official APKs retain the existing package and production signing identity, so 2.4.11-miffan.1 can be upgraded in place without changing the app's data or database compatibility policy. Nightly workflow artifacts use me.ayuilos.miffan.app.nightly and an ephemeral CI debug signer; they install separately and cannot replace an official build. Nightlies are disposable test artifacts, and the signer may change between runs, so in-place upgrades between Nightlies are not guaranteed.

Security and privacy notes

Miffan is a client: prompts, attachments, and tool data are sent to the model, search, speech, MCP, sync, or other endpoints you choose. Review the privacy policy and pricing of every service you configure. The repository's data-flow disclosures are in PRIVACY.md.

Skills, MCP servers, local tools, and workspaces can access external services or device data within their granted scope. Install trusted Skills, inspect tool requests, and enable only the capabilities an assistant needs.

Read SECURITY.md before enabling LAN access, running workspace commands, or handling sensitive credentials.

Build from source

The project uses Kotlin, Jetpack Compose, Material 3, Gradle, and Java 17.

git clone https://github.com/Ayuilos/Miffan.git
cd Miffan
./gradlew assembleDebug

Useful verification commands:

./gradlew test
./gradlew lint

app/google-services.json is optional. Builds without an authorized configuration keep Firebase analytics and crash reporting disabled. Production signing and release steps are documented in docs/releasing.md.

Character animation lab

Open Settings → About, long-press the version entry, and select 蓝色大肥鱼 in the debug page. Preview all eight expressions, pause or replay them, switch light/dark colors and reduced motion, test 28–280 dp sizes, or scrub through the action timeline. These controls affect only the preview.

The whale uses native Compose paths and a foreground animation clock. Unused legacy WebP animation atlases have been removed from the APK. See the character architecture and visual validation notes for implementation and test evidence.

Repository modules

Project history and attribution

Miffan began as a community fork of RikkaHub and continues to incorporate selected upstream improvements. It is now maintained as an independent application with its own product identity, character system, package name, signing certificate, release line, and feature development. RikkaHub is treated as a selective code input rather than a product-version source; the review and provenance rules are documented in the upstream synchronization policy.

Upstream copyright and attribution are retained as required by the license. Miffan is not an official RikkaHub release.

Contributing

Issues and pull requests are welcome. For substantial changes, open an issue first so the product direction and implementation scope can be discussed. When reporting a bug, include the Miffan version, Android version, provider type, and reproducible steps, but remove API keys and private conversation content.

Before opening an issue, please read the Issue guidelines. Before submitting a pull request, please read the contribution guidelines.

License

Miffan is released under the GNU Affero General Public License v3.0.