Product Overview

About DOST

The technical architecture, security principles, and agent execution pipelines powering the DOST assistant.

Core Mission
Bridges conversational AI models with your local OS and cloud apps.

Standard LLMs are locked within conversation boxes. DOST is built for execution. It bridges the gap between natural language prompts and functional OS APIs (volume/brightness, app launchers, file explorers) and cloud services.

By adopting the open Model Context Protocol (MCP), DOST separates the AI model reasoning loop from action execution, enabling secure tool connectivity across any compatible environment.

Security & Privacy
Strict sandbox controls and localized authorization caches.

All desktop tools run locally on your system using standard input/output (stdio) transports, preventing external data leaks. Sensitive command executions are restricted by whitelists and path sanitizers.

Cloud integrations (Google Workspace, Spotify) leverage secure OAuth 2.0. Access credentials are cached inside local JSON files or Valkey/Redis databases, keeping authentication details completely private.

System Architecture
DOST consists of three primary, decoupled components:

1. Desktop Client / App

The Electron-based desktop interface managing active user sessions, prompt histories, local database records, and launching local MCP servers.

Active Workspace (Electron)

2. Next.js Web Dashboard

Handles user enrollment signup, Google and Spotify OAuth callbacks, and generates secure API keys for remote MCP server connection.

Port 3000 (Local Dev)

3. Local & Remote Servers

Stdio daemon (local utilities) and uvicorn FastAPI gateway (exposing stock, weather, calendar, Spotify, and mail tools behind API Key validation).

Stdio & Port 8000
Hybrid ToolRAG Search
Prevents context bloating by indexing schemas dynamically.

DOST utilizes a hybrid search strategy to locate relevant tools. It combines TF-IDF/BM25 text searches with vector embeddings, filtering tool schemas dynamically. The agent receives only the top matching tools for vague queries, keeping the prompt small and cost-effective.

Context Slicing & Checkpoints
Auto-compresses history when token counts spike.

The chat window monitors prompt token sizes. When they exceed VITE_SUMMARY_TRIGGER_TOKENS, the client initiates a background summarization checkpoint, compressing older conversations while retaining current context parameters. Height-compensation scrolls prevent UI jumping.

MIT Licensed

100% Client-Owned Operations

DOST is completely open source and extensible. You can add your own custom servers, run local models, and connect your favorite development editors (such as Claude Desktop and Cursor) with zero lock-in.