SQLite MCP server: setup for Claude & Cursor (2026)

SQLite is the easiest database to wire into an AI client — it is just a file. An MCP server lets the model inspect tables, run queries and help analyse the data. Here are the two clean routes for Claude Desktop and Cursor.

Option A — DBHub (recommended, no Python)

DBHub is a universal database server; for SQLite you just pass a file DSN:

{
  "mcpServers": {
    "sqlite": {
      "command": "npx",
      "args": [
        "-y", "@bytebase/dbhub",
        "--transport", "stdio",
        "--readonly",
        "--dsn", "sqlite:///Users/you/data/app.db"
      ]
    }
  }
}

Same server you would use for MySQL or Postgres — one tool to learn. See MySQL MCP setup for the same pattern.

Option B — Python SQLite server

If you prefer the dedicated Python server, run it with uvx:

{
  "mcpServers": {
    "sqlite": {
      "command": "uvx",
      "args": ["mcp-server-sqlite", "--db-path", "/Users/you/data/app.db"]
    }
  }
}

This needs Python 3.10+ and uv installed.

Restart the client after editing; see MCP config file location.

Verify

Ask: "List the tables in my SQLite database and show 5 rows from users." Real data confirms the connection.

Common problems

• unable to open database file — wrong path, or the file does not exist. Use an absolute path; note the sqlite:/// triple slash for absolute paths in DBHub.
• Locked database — another process holds a write lock; read-only mode avoids most of this.
• uvx: command not found — install uv (Option B), or use Option A which only needs Node.
• No tools in client — see Cursor MCP not working.

Going further

Keep it read-only by default — see MCP security best practices. Comparing engines? Read best MCP server for a database, or browse the databases category.