5ire

Provides a single chat interface that abstracts 12+ AI providers (OpenAI, Anthropic, Google, Mistral, Grok, DeepSeek, Ollama, Perplexity, Doubao, etc.) through a provider-agnostic chat service base architecture. Implements streaming responses via provider-specific SDK integrations, with per-conversation model and parameter configuration. Uses Zustand for state management and React 18.3.1 for real-time message rendering with token counting per provider's native implementation.

Integrates the Model Context Protocol (MCP) via three transport mechanisms: StdioTransport for local processes, SSETransport for HTTP server-sent events, and StreamableHTTPTransport for streaming HTTP. Manages MCP server lifecycle (startup, shutdown, reconnection) in the Electron main process, exposes tool schemas to the chat system, and routes tool execution requests through the MCP protocol with approval policies. Stores MCP server configurations in SQLite for persistence across sessions.

Implements a chat input editor with model and parameter controls (temperature, max_tokens, top_p, etc.) accessible per-conversation. Uses a text input component with support for multi-line input and keyboard shortcuts (Shift+Enter for newline, Enter to send). Provides a parameter panel with sliders and input fields for model-specific settings. Stores parameter configurations per conversation in SQLite, enabling different settings for different conversations. Integrates with the chat service to send prompts with the selected model and parameters.

Implements a document ingestion pipeline that processes PDF, DOCX, XLSX, and TXT files into embeddings. Extracts text from each format using format-specific parsers (PDF.js for PDFs, docx library for Word docs, xlsx library for spreadsheets). Chunks extracted text into overlapping segments (default chunk size ~512 tokens with overlap). Generates embeddings using bge-m3 model via @xenova/transformers for client-side inference. Stores embeddings in LanceDB with document metadata (filename, upload_date, file_size) in SQLite. Provides progress tracking for long-running ingestion operations.

Implements a local-first document ingestion pipeline that processes PDFs, DOCX, XLSX, and TXT files into embeddings using bge-m3 model (@xenova/transformers for client-side inference). Stores embeddings in LanceDB vector database with document metadata in SQLite. Provides semantic search across the knowledge base with citation tracking, integrating search results into chat context as RAG (Retrieval-Augmented Generation). Uses PGLite for optional in-process vector operations.

Implements a provider management system that dynamically discovers available models from each provider's API (e.g., OpenAI's list_models endpoint). Stores provider configurations and API keys in Electron Store with encryption at rest. Supports custom provider configuration for self-hosted or alternative endpoints. Maintains a provider registry with per-provider token counting strategies and model metadata (context window, pricing). Allows runtime provider switching without application restart.

Implements a tool execution system where MCP tools are exposed to the AI model, but execution is gated by configurable approval policies (auto-approve, user-confirm, deny). Tool invocation requests from the model are intercepted in the chat service, validated against the approval policy, and either executed immediately or presented to the user for confirmation. Execution happens in the Electron main process with access to the MCP server, maintaining a tool execution audit log in SQLite.

Built on Electron 31.7.1 with a three-process model: Main Process (Node.js) manages application lifecycle and system integration, Renderer Process (Chromium + React 18.3.1) handles UI rendering, and Preload Script provides sandboxed context bridge for secure IPC. Uses Fluent UI components for native OS appearance (Windows, macOS, Linux). Implements persistent state management with Zustand for UI state and SQLite (better-sqlite3) for application data, with optional Supabase sync for cloud backup.

Implements per-provider token counting strategies that abstract provider-specific token counting APIs (OpenAI's tiktoken, Anthropic's token counter, etc.). Tracks token usage per message, conversation, and provider in SQLite. Calculates estimated costs based on provider pricing models (input/output token rates). Displays real-time token usage in the chat UI and provides usage analytics dashboard. Handles token counting for both prompts and completions with fallback estimation for providers without native token counting.

Implements a localization system using i18n (internationalization) with JSON-based translation files stored in public/locales/{language}/translation.json. Supports English and Chinese (Simplified) with extensible architecture for additional languages. Uses React context or Zustand for language state management, enabling runtime language switching without application restart. Translation keys are namespaced by feature (chat, tools, settings, etc.) for maintainability.

Implements a conversation management system that stores chat history in SQLite with per-conversation metadata (title, created_at, updated_at, provider, model). Supports conversation creation, deletion, renaming, and archiving. Maintains message history with role (user/assistant), content, and metadata (tokens, provider, model). Provides conversation search and filtering by date, provider, or model. Integrates with Zustand state management for UI state (current conversation, message list) and SQLite for persistent storage.

Implements streaming message rendering in React that displays AI responses character-by-character as they arrive from the provider. Uses React state updates to append chunks to the message content, with optimized re-rendering to avoid jank. Supports markdown rendering with syntax highlighting for code blocks. Handles message metadata display (tokens, model, provider, timestamp). Integrates with the chat service to receive streaming chunks and update UI in real-time.