Qurio - Multi-Provider AI Knowledge Workspace

Not another chat UI. Qurio is a research-grade cockpit for real work.

Qurio is a high-velocity AI knowledge workspace built for teams that demand more than basic chat. It supports generic OpenAI-compatible providers, Gemini, SiliconFlow, Kimi (Moonshot), MiniMax, GLM (Zhipu), NVIDIA NIM, and more. Highlights include Deep Research for complex tasks, Custom Agents for specialized workflows, rich tool orchestration (MCP + HTTP), and structured reasoning views with export-ready outputs. The stack is designed around separation of concerns: lightweight models for routing/planning, stronger models for answers, and structured storage for memory and documents.

In one line: plan, tool, and deliver from a single workspace without switching apps.

Screenshots

Home

Conversation + Tool Calls

Deep Research Workflow

MCP Tool Groups

Why Qurio

• Multi-Provider Powerhouse: Connect to almost any model (Gemini, OpenAI, Claude via compatible endpoints, SiliconFlow, Kimi, MiniMax, GLM, NIM, etc.) with dual model slots (Lite + Default).
• Tool Orchestration Beyond LLM Chat: Built-in MCP + HTTP tooling, tool grouping, and per-tool enable/disable controls so agents only see what you want them to use.
• Deep Research, Not Just Search: Dedicated research workflow that plans, executes, and reports with structured steps, citations, and export-ready outputs.
• Agent System That Scales: Per-agent prompts, model preferences, and toolsets with clean switching and workspace binding.
• Document Intelligence: Parse, embed, and retrieve across uploaded documents with hybrid search (vector + keyword) for grounded answers.
• Long-Term Memory That Stays Useful: Domain-indexed summaries with Lite routing so only relevant preferences/background get injected.
• Premium UX: Streaming, rich Markdown/Code rendering, inline images, reasoning visibility, and thoughtful UI for research/think/tool panels.
• Production-Grade: Light/Dark/System themes, secure API key handling, i18n (en/zh), responsive design, Supabase persistence.

Design Philosophy

• Route first, retrieve second: Lite models decide if memory/docs are needed; only then inject relevant snippets.
• Separation of models: Fast models handle planning/routing; stronger models handle final answers.
• Structured memory over keyword lists: Domain-indexed summaries keep personalization usable and controllable.
• Hybrid retrieval: Vector + keyword search balances recall with precision for real-world docs.
• Tool governance: MCP groups, HTTP tools, and per-tool visibility keep agents safe and focused.

Highlights

📱 Mobile PWA Support

• Installable: Add Qurio to your home screen on iOS/Android for an app-like experience.
• Responsive UX: Optimized layouts for mobile, tablets, and desktops with touch-friendly controls.
• Seamless Sessions: Stay signed in and continue where you left off with persistent conversations.

🔬 Academic Research Reports

• Automated Research Pipeline: Deep Research mode breaks down complex research goals into structured plans, conducts iterative web searches, and synthesizes comprehensive academic-style reports with proper citations.
• Source Tracking: All claims are backed by timestamped sources with direct links, ensuring research integrity and easy verification.
• Export-Ready: Generate polished reports in Markdown format, ready for academic or professional use.

🤖 Intelligent Agent System

• Custom Agents: Create specialized AI personas with unique system prompts, model preferences, and personality traits (tone, warmth, enthusiasm).
• Automatic Agent Selection: The system can intelligently select the best agent for your task based on context, or you can manually choose from your agent library.
• Multi-Model Support: Each agent can have dual model slots (Default + Lite) from different providers, optimizing for quality vs. speed.
• Space-Agent Binding: Assign specific agents to workspaces for consistent behavior across related conversations.

🧰 MCP + HTTP Tooling

• MCP Tool Groups: Connect external MCP servers (SSE or Streamable HTTP) and manage tools as groups.
• Per-Tool Enable/Disable: Keep tools saved while selectively hiding them from agents.
• HTTP API Tools: Define custom tools by URL, method, and JSON params; schemas are auto-generated from templates.
• Interactive Forms: Tools can request structured user input inline before execution.

🧾 Interactive Forms

• Dynamic Form Generation: AI agents can request structured input through beautifully rendered forms with various field types (text, number, select, textarea, etc.).
• Real-time Validation: Form inputs are validated before submission, ensuring data quality.
• Seamless UX: Forms appear inline in the conversation flow, maintaining context while collecting detailed information.

Document Embeddings & Hybrid Search

• Upload & Parse: Add documents to a space; text is parsed and chunked automatically.
• Hybrid Retrieval: Combines embeddings with keyword search for stronger recall.
• Context Controls: Only selected documents are retrieved and injected per conversation.

Long-Term Memory (Domain Routing)

• Domain Index + Summary: Memory is stored as domain cards with concise summaries.
• Lite Gating: A lightweight model decides if memory is needed and which domains to inject.
• Selective Injection: Only hit domains are appended, keeping prompts tight and relevant.

🌐 Extended Provider Support

Out-of-the-box support for:

• Google Gemini (Native SDK with Thinking mode)
• SiliconFlow (DeepSeek-R1, Qwen, etc.)
• Kimi (Moonshot)
• MiniMax
• GLM (Zhipu AI)
• ModelScope
• NVIDIA NIM
• OpenAI & Compatibles (Any standard endpoint)

🧩 Production-Grade Features

• Question Timeline: Interactive history cards grouped by time (Today, Yesterday, Past 7 Days) with instant search.
• Security & Privacy: API keys stored securely in Supabase or kept in-memory (never in LocalStorage for sensitive keys).
• Personalization: Adjust font sizes, interface language (en/zh), AI response styles, and theme (Light/Dark/System).
• Media Support: Drag & drop images, paste screenshots, and render inline media with vision-capable models.

Tech Stack

• Frontend: React 19, TanStack Router v1, Zustand (State), Tailwind CSS v4, Lucide Icons — componentized UI with predictable state.
• Build: RSBuild + Bun — fast dev loops and reproducible builds.
• Data: Supabase (PostgreSQL) — structured persistence for spaces, messages, documents, and long-term memory.
• AI Integration: Custom adapter layer that normalizes providers and enforces model routing (Lite vs Default).
• Retrieval: Embeddings + keyword hybrid search to reduce misses and hallucinations.
• Rendering: react-markdown, remark-gfm, react-syntax-highlighter for reliable output UX.

Getting Started

1. Prerequisites

   • Bun 1.3+
   • A Supabase project (for persistence)
   • Tavily API Key (essential for Web Search & Deep Research)

2. Install

   git clone <your-repo-url>
   cd qurio
   bun install
   cd backend
   bun install

3. Environment Copy .env.example to .env and configure your keys. Key variables include:

   • PUBLIC_SUPABASE_URL, PUBLIC_SUPABASE_KEY
   • PUBLIC_TAVILY_API_KEY (Required for Search/Deep Research)

   Provider Keys (Optional depending on usage):

   • PUBLIC_GOOGLE_API_KEY
   • PUBLIC_OPENAI_API_KEY & PUBLIC_OPENAI_BASE_URL
   • PUBLIC_SILICONFLOW_API_KEY
   • PUBLIC_KIMI_API_KEY
   • PUBLIC_MINIMAX_API_KEY
   • PUBLIC_GLM_API_KEY
   • PUBLIC_MODELSCOPE_API_KEY
   • PUBLIC_NVIDIA_API_KEY

4. Database Setup In your Supabase dashboard, open the SQL Editor and run the contents of supabase/schema.sql. This creates the necessary tables (spaces, conversations, messages, settings, etc.) and Row Level Security (RLS) policies.

5. Run (Frontend + Backend)

   # One command (frontend + backend)
   bun run dev:full

   Frontend: http://localhost:5173
   Backend: http://localhost:3001

   # Frontend only
   bun run dev
   
   # Backend only
   bun run dev:backend

6. Build

   bun run build

Usage Tips

• Settings: Click the gear icon to set up providers. You can verify your Supabase connection and input API keys directly in the UI if you didn't set them in .env.
• Deep Research: Navigate to the "Deep Research" tab (Microscope icon) to start a comprehensive research task. This mode automatically browses the web and aggregates information.
• Agents: Use the "Agents" menu to create specialized personas. Assign a "Coding God" agent to use a high-reasoning model like DeepSeek-R1 or o1, and a "Casual Chat" agent to use Gemini Flash.
• Search: The "Web Search" toggle uses Tavily to fetch real-time info. It's automatically invoked by Deep Research but can be toggled manually in normal chat.
• Shortcuts:
  • Enter to send, Shift+Enter for newline.
  • Cmd/Ctrl + K (if implemented) or UI buttons to switch agents/models quickly.
• Styling: Go to Settings -> Appearance to toggle Dark Mode or adjust Font Size.

Project Structure

• /src/lib/providers.js: Core logic for adapting different AI APIs.
• /src/components/chat: Chat interface components (Input, Bubble, Header).
• /src/views: Top-level route views (Home, Conversation, DeepResearch, Spaces).
• /supabase: SQL migrations and schema definitions.

Contributing

Pull requests are welcome for improvements (UI polish, new providers, bug fixes). Please open an issue for major feature requests.

License

Non-commercial: Provided for personal and educational use. See LICENSE.