Cove

Cove is a multi-agent AI research assistant that performs query-focused literature reviews, web searches, and citation verification. It retrieves academic papers (arXiv, OpenAlex) and web pages (Tavily), extracts factual claims, and verifies them directly against source passages to prevent hallucinations.

Cove runs as both a web application and a Model Context Protocol (MCP) server.

---

Architecture

Cove coordinates multiple specialized agents using a ReAct (Reasoning and Action) orchestration loop.

System Diagram

graph TD
    User([User Query / MCP Call]) --> InputGuard[Input Guardrails]
    InputGuard --> Orchestrator{Orchestrator <br/> ReAct Loop}
    
    Orchestrator -->|1. Search| SearchAgent[Unified Search]
    SearchAgent --> Academic[arXiv / OpenAlex APIs]
    SearchAgent --> Web[Tavily Search API]
    Academic & Web --> Dedupe[Deduplicate & Rank]
    Dedupe --> Relevance[Batch Relevance Check <br/> Llama 3.1 8B]
    Relevance -->|Top Chunks| VectorDB[(pgvector DB)]
    
    Orchestrator -->|2. Summarize| Summarizer[Summarizer Agent <br/> Extract Claims]
    Summarizer --> Chunks[Read Passages / PDF Chunks]
    
    Orchestrator -->|3. Verify| Verifier[Verifier Agent <br/> Llama 3.3 70B]
    Verifier --> CrossCheck[Cross-Check Claims vs Sources]
    
    Orchestrator -->|4. Finish| Reporter[Report Writer]
    Reporter --> Report[JSON Synthesis]
    Report --> OutputGuard[Output Guardrails]
    OutputGuard --> Out([React UI / MCP Markdown])

Loading

Execution Flow

1. Input Guard: The query is validated for safety via a Groq-hosted safety model.
2. ReAct Loop (Orchestrator): Evaluates what info is missing, selecting tools dynamically:
   • search_query: Searches databases. Runs local bge-small-en-v1.5 embeddings to rank matches, filters the top 8 candidates using a batched Llama 3.1 relevance audit, and indexes contents into pgvector (if deep research is enabled).
   • summarize_sources: Scrapes pages/PDFs and extracts factual claims.
   • verify_claims: Evaluates all claims in a single batch call per paper using Llama 3.3 70B to assign confidence states (Supported, Contradicted, etc.).
3. Report Generation: Synthesizes verified claims into a markdown report containing:
   • Contradiction Detection: Conflicting viewpoints across papers.
   • Research Gaps: Missing information in the literature.
   • Evidence Graph: A citation matrix mapping claims to source URLs.

---

MCP Server Integration

Cove exposes its research pipeline as an MCP server. You can integrate it with Claude Desktop or any MCP-compatible client.

MCP Tools

Tool Name	Description	Key Arguments
run_research	Runs the full research pipeline and outputs a cited Markdown survey.	query (string), deepResearch (bool)
unified_search	Searches academic + web databases, deduplicates, and semantically ranks.	query (string)
scrape_url	Downloads any web page or PDF and extracts clean readable text.	url (string)
store_document	Scrapes, chunks, embeds, and indexes a URL into the local vector DB.	url (string)
query_knowledge_base	Queries the local pgvector store for matching text chunks.	query (string), k (int)

Claude Desktop Configuration

Add this to your Claude Desktop configuration file (typically C:\Users\<username>\AppData\Roaming\Claude\claude_desktop_config.json on Windows or ~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "cove": {
      "command": "node",
      "args": ["C:\\...\\cove\\backend\\mcp-server.js"]
    }
  }
}

Cove automatically loads environment variables (including API keys and DB URLs) from the local backend/.env file relative to the script path. You do not need to expose your keys in the Claude configuration file.

---

Local Development Setup

Prerequisites

• Node.js 18+
• PostgreSQL with pgvector extension (optional, required for deep research RAG)

Setup Steps

1. Get API Keys:

   • Groq: API key from console.groq.com
   • Gemini: API key from aistudio.google.com/app/api-keys
   • Tavily: Search API key from tavily.com
   • Firebase: Set up a Firestore project. Download your admin credentials to backend/serviceAccountKey.json and configure Google authentication.

2. Backend Setup:

   cd backend
   npm install

   Create a backend/.env file:

   GROQ_API_KEY1=your_groq_api_key
   GROQ_API_KEY2=optional_backup_key_1
   GEMINI_API_KEY1=your_gemini_api_key
   GEMINI_API_KEY2=optional_backup_key_1
   GEMINI_API_KEY3=optional_backup_key_2
   TAVILY_API_KEY=your_tavily_key
   DATABASE_URL=your_postgres_db_url
   PORT=8000

   Start the backend server:

   npm run dev

3. Frontend Setup:

   cd ../frontend
   npm install

   Create a frontend/.env file:

   REACT_APP_API_URL=http://localhost:8000
   # Paste your Firebase Client configuration:
   REACT_APP_FIREBASE_API_KEY=...
   REACT_APP_FIREBASE_AUTH_DOMAIN=...
   REACT_APP_FIREBASE_PROJECT_ID=...
   REACT_APP_FIREBASE_STORAGE_BUCKET=...
   REACT_APP_FIREBASE_MESSAGING_SENDER_ID=...
   REACT_APP_FIREBASE_APP_ID=...

   Start the frontend React app:

   npm run dev

---

Codebase Map

cove/
├── backend/
│   ├── agents/            # Orchestrator ReAct loop & report generator
│   ├── db/                # Neon/Drizzle schema definitions & loaders
│   ├── guardrails/        # Content safety policy checks
│   ├── ranking/           # Local ONNX semantic ranking algorithms
│   ├── retrieval/         # Academic/web crawlers & vector store integrations
│   ├── utils/             # Rotated Groq/Gemini API clients
│   ├── verification/      # Citation claim checking engines
│   ├── mcp-server.js      # Stdio MCP entrypoint
│   └── index.js           # HTTP/SSE Express server entrypoint
└── frontend/
    ├── public/            
    └── src/
        ├── components/    
        ├── hooks/         
        └── theme.js       # Design tokens & styles