db-semantic-mcp

English | 简体中文

A multi-backend MCP server for AI coding agents — supports both PostgreSQL and SQL Server.

Exposes your database schema — table names, column types, comments, and bounded sample data — as MCP tools. Includes semantic search powered by any OpenAI-compatible LLM, enriched by a user-authored semantic layer document.

No SQL execution. Read-only. No vector database required.

Backends

The backend is auto-detected from DATABASE_URL. Everything else works the same.

Features

• list_tables — discover all tables with comments
• describe_table — inspect column names, types, nullability, and comments
• sample_data — fetch example rows from any table
• search_schema — semantic keyword search across tables and columns using LLM

Install

# PostgreSQL only
pip install db-semantic-mcp

# With SQL Server support
pip install "db-semantic-mcp[sqlserver]"

Requires Python 3.11+.

Quick Start

# PostgreSQL
export DATABASE_URL="postgresql://user:pass@localhost:5432/mydb"

# SQL Server (Kingdee ERP or any MSSQL instance)
export DATABASE_URL="sqlserver://user:pass@host:1433?database=mydb&encrypt=disable"

export LLM_API_KEY="sk-..."          # required only for search_schema
pg-semantic-mcp

Configuration

You can also use a .env file in the working directory.

sample_data Security Boundary

sample_data is read-only, but it is not metadata-only: it can expose actual business data from the connected database. Treat it as a small data-plane tool.

The tool applies these response controls before returning rows to the MCP client:

• limit must be greater than 0 and is hard-capped by SAMPLE_DATA_MAX_ROWS.
• The response is reduced until its serialized size is within SAMPLE_DATA_MAX_BYTES.
• SAMPLE_DATA_ALLOW_COLUMNS limits returned columns when set.
• SAMPLE_DATA_DENY_COLUMNS omits matching columns and takes precedence over allow rules.
• SAMPLE_DATA_REDACT_COLUMNS masks matching values with [REDACTED].

Column policies use case-insensitive glob patterns. For example:

SAMPLE_DATA_ALLOW_COLUMNS="id,name,email,created_at"
SAMPLE_DATA_DENY_COLUMNS="*password*,*token*"
SAMPLE_DATA_REDACT_COLUMNS="*email*,*phone*,*secret*"

The response shape is intentionally short to save model context:

{
  "rows": [
    {"id": 1, "email": "[REDACTED]"}
  ],
  "_meta": {
    "table": "public.customers",
    "returned": 1,
    "truncated": false
  }
}

Register with OpenCode

Add to your opencode.jsonc:

{
  "mcp": {
    "pg-data": {
      "type": "local",
      "command": "pg-semantic-mcp",
      "environment": {
        "DATABASE_URL": "postgresql://user:pass@host:5432/dbname",
        "SEMANTIC_FILE": "/path/to/SCHEMA.md",
        "LLM_API_KEY": "sk-..."
      }
    }
  }
}

Same config format works for Claude Code, Cursor, and any MCP-compatible agent.

Semantic Layer

Create a SCHEMA.md file describing your database — naming conventions, business term mappings, design decisions. See SCHEMA.md.example for a template.

This document is loaded at startup and included in the search_schema LLM prompt. It is the main way to teach the agent about your specific domain.

Compatible LLMs

search_schema calls any OpenAI-compatible endpoint:

• OpenAI (gpt-4o-mini, gpt-4o, …)
• DeepSeek (deepseek-v4, set LLM_BASE_URL=https://api.deepseek.com/v1)
• Anthropic via proxy
• Local models via Ollama or LM Studio

License

MIT