localmind

LocalMind is a Flutter mobile chat client for talking to AI models without sending your conversations to anyone else's infrastructure. It speaks directly to user-configured servers — on-device inference, local Ollama or LM Studio instances, or any OpenAI-compatible API such as OpenRouter — with no middleware, account, telemetry, or subscription in between.

Why it matters

• True privacy floor. Zero analytics, no first-party servers, and no fallback proxy. Every byte leaves the device only if the user points the app at a server they trust, and conversation history is stored in encrypted Hive locally.
• Runs models on the phone. GGUF and LiteRT checkpoints (Gemma, Qwen, DeepSeek R1 and friends) execute directly on iOS or Android via llamadart and flutter_gemma, so the same app works fully offline once a model is downloaded.
• MCP for tool use. The app ships a Model Context Protocol client, so a running model can be wired to local or remote MCP servers for file access, shell actions, or web APIs — toggled per server.
• Multi-server by design. Multiple inference endpoints can coexist with live health monitoring, live model swapping mid-conversation, streaming SSE, expandable reasoning traces (for DeepSeek R1-style models), markdown + syntax highlighting, voice I/O, and image attachments for vision models.

How it works

The codebase is a clean Flutter project (lib/bootstrap, lib/core, lib/features, lib/services, lib/l10n) with Riverpod for state and go_router for navigation. The features folder is organised as vertical slices — chat, conversations, models, on_device, mcp, personas, servers, settings, saved_messages, cloud_sync, lm_studio_catalog, stt, tts, voice_mode, sidebar, onboarding — so each capability owns its own UI and controllers.

Persistence is dual-layered: ObjectBox (objectbox + objectbox_flutter_libs) holds the relational chat history, while Hive (with flutter_secure_storage for secrets) holds lightweight key-value preferences. Networking is plain dio to whatever OpenAI-compatible endpoint the user configured, with url_launcher for external links and aws_signature_v4 for the optional S3 cloud-sync target that ships with end-to-end encryption.

Caveats

• On-device inference is hardware-bound. A 7B GGUF on a mid-tier Android will be slow; phones older than ~2022 may not run useful models at all. The cloud / local-server paths are the realistic default for most users.
• Single-user app. There is no notion of shared accounts or team workspaces — this is a personal client, not a SaaS frontend.
• MCP tools are a power feature. A misconfigured MCP server can give a model shell access; treat endpoint URLs with the same care you would give any LLM agent tool.
• iOS App Store status is unclear. The repo publishes source only; distribution is via flutter run or sideload at present.

Deployment notes

git clone https://github.com/abdulmominsakib/localmind.git
cd localmind
flutter pub get
flutter run

For on-device inference you also need a GGUF or LiteRT model — either pull one through the in-app Model Manager (HuggingFace integration) or drop a file into the app's documents directory. For local-server mode, run Ollama or LM Studio on the same network and add the endpoint under Settings → Servers; OpenRouter or any custom OpenAI-compatible host works the same way.

Integration tip: if you curate a directory like Astro/Grove for self-hosted or privacy-first tooling, LocalMind pairs naturally with any project that exposes an OpenAI-compatible API — Ollama, LM Studio, vLLM, llama.cpp's server mode — and is one of the cleanest reference apps for showcasing on-device LLM UX in Flutter.