db-insight MCP Server

db-insight

Local-first SQL insight CLI and MCP server for Postgres and SQLite.

Positioning: connect a read-only replica, analytics database, or staging database. Keep credentials on your machine while the tool discovers schema, generates safe SQL, previews it for approval, runs the query, and summarizes the result.

Install

python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

Configure

Create .env:

Postgres:

DATABASE_URL=postgresql://readonly:password@localhost:5432/app_db
DB_INSIGHT_MODEL=gemma3:latest
DB_INSIGHT_OLLAMA_URL=http://localhost:11434

SQLite:

DATABASE_URL=sqlite:///absolute/path/to/app.sqlite
DB_INSIGHT_MODEL=gemma3:latest
DB_INSIGHT_OLLAMA_URL=http://localhost:11434

If your provider gives a URL with special characters, wrap it in quotes:

DATABASE_URL="postgresql://readonly:[email protected]:5432/app_db"

DB_INSIGHT_MODEL is configurable so you can use the best Gemma/Ollama tag available on your machine.

CLI

db-insight connect
db-insight schema --refresh
db-insight ask "Why did revenue drop last week?"

By default, ask uses the same MCP-style flow as the server:

user question
→ LLM plans tool calls
→ discover_schema / schema memory
→ generate_safe_sql
→ preview SQL for approval
→ run_approved_sql
→ summarize_results

The user-facing CLI previews generated SQL and asks before running it.

db-insight schema --refresh stores a local schema memory snapshot in .db-insight/schema_memory.json. Later questions reuse that schema picture so the model does not need to rediscover the same database structure every time. The memory contains metadata only, not table rows.

MCP stdio server

For Android health-ring chat architecture, see Android Health Chat Implementation.

db-insight mcp

This exposes safe local tools over stdio for MCP clients:

• plan_tool_calls
• discover_schema
• refresh_schema_memory
• schema_memory_status
• catalog_overview
• inspect_table
• explain_sql
• generate_safe_sql
• run_approved_sql
• ask_database

In a full MCP client, the LLM decides which tools to call. ask_database is a convenience tool that performs that decision loop inside this local server while still requiring explicit approval before execution.

Remote MCP server

For team use, run the same server over HTTP:

DB_INSIGHT_MCP_HOST=0.0.0.0 \
DB_INSIGHT_MCP_PORT=8000 \
db-insight mcp --transport streamable-http

The HTTP MCP endpoint is:

http://your-host:8000/mcp

Put it behind your existing auth layer, VPN, or private network. Do not expose a database-backed MCP server directly to the public internet.

Docker

Build the image:

docker build -t db-insight:latest .

Or use the published image:

docker pull ghcr.io/enclavex-labs/db-insight:latest

Install Gemma on the host Docker server:

ollama pull gemma3:latest

Configure your MCP client to launch the container over stdio.

Postgres:

{
  "mcpServers": {
    "db-insight": {
      "command": "docker",
      "args": [
        "run",
        "-i",
        "--rm",
        "--add-host=host.docker.internal:host-gateway",
        "-e",
        "DATABASE_URL",
        "-e",
        "DB_INSIGHT_MODEL",
        "-e",
        "DB_INSIGHT_OLLAMA_URL",
        "ghcr.io/enclavex-labs/db-insight:latest"
      ],
      "env": {
        "DATABASE_URL": "postgresql://readonly:[email protected]:5432/app_db",
        "DB_INSIGHT_MODEL": "gemma3:latest",
        "DB_INSIGHT_OLLAMA_URL": "http://host.docker.internal:11434"
      }
    }
  }
}

SQLite:

{
  "mcpServers": {
    "db-insight": {
      "command": "docker",
      "args": [
        "run",
        "-i",
        "--rm",
        "--add-host=host.docker.internal:host-gateway",
        "-v",
        "/absolute/path/to/data:/data:ro",
        "-e",
        "DATABASE_URL",
        "-e",
        "DB_INSIGHT_MODEL",
        "-e",
        "DB_INSIGHT_OLLAMA_URL",
        "ghcr.io/enclavex-labs/db-insight:latest"
      ],
      "env": {
        "DATABASE_URL": "sqlite:////data/app.sqlite",
        "DB_INSIGHT_MODEL": "gemma3:latest",
        "DB_INSIGHT_OLLAMA_URL": "http://host.docker.internal:11434"
      }
    }
  }
}

For VS Code/Copilot, use the same body under servers instead of mcpServers.

Each user fills DATABASE_URL with their own Postgres connection string or SQLite file URL. For SQLite in Docker, mount the folder containing the database to /data and use four slashes: sqlite:////data/app.sqlite.

For Postgres, if DATABASE_URL uses localhost or 127.0.0.1, the Docker image remaps it to host.docker.internal automatically.

If you need a long-running shared HTTP endpoint instead, run:

docker run --rm -p 8000:8000 \
  --add-host=host.docker.internal:host-gateway \
  -e DATABASE_URL=postgresql://readonly:[email protected]:5432/app_db \
  -e DB_INSIGHT_OLLAMA_URL=http://host.docker.internal:11434 \
  ghcr.io/enclavex-labs/db-insight:latest db-insight mcp --transport streamable-http