Msty

Provides a single conversation interface that abstracts away differences between local models (running via Ollama, LM Studio, or similar) and remote API-based models (OpenAI, Anthropic, etc.). The application maintains a model registry that maps provider-specific connection details and authentication to a normalized chat protocol, allowing users to switch between model backends without changing their interaction pattern or conversation history structure.

Manages the lifecycle and resource allocation for running large language models directly on the user's machine by interfacing with local inference engines like Ollama or LM Studio. The application handles model downloading, GPU/CPU resource allocation, context window management, and inference parameter tuning without requiring users to interact with command-line tools or manage system resources manually.

Delivers a responsive, native-feeling user interface across Windows, macOS, and Linux using a modern desktop framework (likely Electron or similar). The application prioritizes performance and responsiveness, with fast model switching, instant conversation loading, and smooth streaming rendering. UI state is managed efficiently to handle long conversation histories without lag.

Maintains stateful conversation threads that preserve full message history, role attribution (user/assistant), and metadata across sessions. The application implements a conversation store that tracks turn-by-turn exchanges, allowing users to reference earlier messages, branch conversations, or resume previous chats. Context is managed at the application level rather than relying on the model to infer conversation state from a single prompt.

Exposes inference parameters (temperature, top_p, max_tokens, repetition_penalty, etc.) through a configuration UI that allows users to adjust model behavior without editing configuration files or API calls. The application translates user-friendly parameter names into provider-specific formats (OpenAI's API parameters vs. Ollama's parameters) and applies them to each inference request, enabling fine-tuning of response creativity, length, and consistency.

Provides a system for saving, organizing, and reusing prompt templates with variable substitution. Users can define templates with placeholders (e.g., {{topic}}, {{language}}) that are filled in at runtime, enabling rapid iteration on prompt engineering and consistent application of refined prompts across multiple conversations. Templates are stored locally and can be organized into categories or collections.

Renders model responses token-by-token as they are generated, providing real-time visual feedback of inference progress. The application handles streaming protocol differences between providers (OpenAI's Server-Sent Events, Anthropic's streaming format, Ollama's streaming output) and displays tokens incrementally in the UI, allowing users to see partial responses and interrupt generation if needed.

Exports conversations in multiple formats (Markdown, JSON, PDF, HTML) for sharing, archiving, or integration with external tools. The application serializes conversation history including metadata (timestamps, model used, parameters) and renders it in format-specific layouts. Export can include or exclude system prompts, metadata, and formatting options.

Securely stores and manages API keys and authentication credentials for multiple model providers in a local credential store. The application encrypts sensitive data at rest, provides a UI for adding/removing/updating credentials, and handles credential injection into API requests transparently. Supports multiple credentials per provider for load balancing or account switching.

Provides a UI for discovering available models from configured providers and adding new providers. The application queries provider APIs or local model registries (Ollama's model list) to populate available models, displays model metadata (context window, parameters, pricing), and guides users through provider-specific setup (API key entry, local server configuration). Supports dynamic provider registration without code changes.

Allows users to define system prompts or role definitions that shape model behavior for entire conversations. System prompts are prepended to each request and persist across turns, enabling consistent persona adoption (e.g., 'You are a Python expert') or instruction sets. The application manages system prompt storage, allows quick switching between predefined roles, and displays the active system prompt in the UI.