MCP Server for AI Agents

This package includes an MCP (Model Context Protocol) server that allows AI agents like Claude to interact with your PGVector database directly.

What is MCP?

Model Context Protocol (MCP) is an open standard that allows AI assistants to connect to external tools and data sources. Think of it as giving AI agents “hands” to interact with your systems.

This is NOT for n8n workflows - n8n uses the PGVector Advanced node directly. MCP is for AI coding assistants and agents.

Who is this for?

Quick Start

1. Install the package

npm install -g n8n-nodes-pgvector-advanced

2. Set environment variables

export PGHOST=localhost
export PGPORT=5432
export PGDATABASE=your_db
export PGUSER=your_user
export PGPASSWORD=your_password

3. Run the MCP server

pgvector-mcp

Or run directly without installing:

npx n8n-nodes-pgvector-advanced

Configure Claude Desktop

Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on Mac):

{
  "mcpServers": {
    "pgvector": {
      "command": "pgvector-mcp",
      "env": {
        "PGHOST": "localhost",
        "PGPORT": "5432",
        "PGDATABASE": "your_db",
        "PGUSER": "your_user",
        "PGPASSWORD": "your_password"
      }
    }
  }
}

Restart Claude Desktop to load the server.

Available Tools

The MCP server provides 5 tools:

pgvector_upsert

Insert or update embeddings in the vector store.

{
  "collection": "documents",
  "externalId": "doc-123",
  "content": "Document text content",
  "embedding": [0.1, 0.2, 0.3, ...],
  "metadata": {"category": "tech", "author": "Jane"}
}

pgvector_query

Search for similar embeddings.

{
  "collection": "documents",
  "queryEmbedding": [0.1, 0.2, ...],
  "topK": 10,
  "distanceMetric": "cosine",
  "metadataFilter": {"category": "tech"}
}

pgvector_delete

Delete embeddings by ID, external ID, or metadata.

{
  "deleteBy": "externalId",
  "collection": "documents",
  "externalIds": ["doc-1", "doc-2"]
}

pgvector_get

Retrieve specific embeddings.

{
  "getBy": "externalId",
  "collection": "documents",
  "externalIds": ["doc-123"]
}

pgvector_admin

Manage schema and indexes.

{
  "operation": "ensureSchema",
  "dimensions": 1536
}

{
  "operation": "createIndex",
  "collection": "documents",
  "indexType": "hnsw",
  "distanceMetric": "cosine"
}

Example Conversation with Claude

Once configured, you can ask Claude things like:

“Store this document in my vector database with the embedding I provide”

“Search for documents similar to this text in the ‘articles’ collection”

“Delete all embeddings with metadata category ‘outdated’”

“Set up the database schema for 1536-dimension embeddings”

Claude will use the MCP tools to perform these operations on your PostgreSQL database.

Docker Setup

If your PostgreSQL is running in Docker, make sure the MCP server can reach it:

# Use host.docker.internal on Mac/Windows
export PGHOST=host.docker.internal

# Or use the container's IP
export PGHOST=$(docker inspect -f '' your-postgres-container)

Troubleshooting

“Connection refused”

“Tool not found in Claude”

“Permission denied”

n8n vs MCP

Feature n8n Node MCP Server
For Workflow automation AI agents
Interface Visual node editor Natural language
Use case Scheduled jobs, triggers Interactive AI chat
Setup Install in n8n Configure in Claude

Use n8n when you want automated workflows with triggers, schedules, and integrations.

Use MCP when you want an AI assistant to interactively work with your vector data.

Next Steps