Database MCP Server

MCP 서버는 AI 클라이언트가 SQL 데이터베이스에 안전하고 구조화된 방식으로 접근할 수 있도록 합니다. 데이터베이스 엔진별로 하나의 설치 가능한 패키지가 제공됩니다. 각 패키지는 동일한 최소한의 두 가지 도구(execute_sql 및 search_objects)를 노출하며, 기본적으로 읽기 전용 모드, 행 제한, 명령문 시간 제한과 같은 보호 장치가 활성화되어 있습니다.

GitHub
5

문서

database-mcp

CI License: MIT npm PyPI crates.io Node >= 22

MCP servers that give AI clients safe, structured access to SQL databases.

One installable package per database engine, in TypeScript (npm), Python (PyPI), and Rust (crates.io). Every package exposes the same minimal two-tool surface, execute_sql and search_objects, with guardrails on by default: read-only mode, row caps, and statement timeouts.

Packages

EngineTypeScript (npm)Python (PyPI)Rust (crates.io)
SQLite@database-mcp/sqlite npmdatabase-mcp-sqlite PyPIdatabase-mcp-rs-sqlite crates.io
libSQL@database-mcp/libsql npmdatabase-mcp-libsql PyPIdatabase-mcp-rs-libsql crates.io
MySQL@database-mcp/mysql npmdatabase-mcp-mysql PyPIdatabase-mcp-rs-mysql crates.io
MariaDB@database-mcp/mariadb npmdatabase-mcp-mariadb PyPIdatabase-mcp-rs-mariadb crates.io
Postgres@database-mcp/postgres npmdatabase-mcp-postgres PyPIdatabase-mcp-rs-postgres crates.io

All three lines are published and pass the same language-agnostic conformance suite against real databases in CI, so behavior is identical regardless of language. Every engine is also listed on the MCP Registry with all three install options: npm, PyPI, and cargo.

Rust crates are named database-mcp-rs-<engine> because the plain names are taken on crates.io; the command each one installs is still database-mcp-<engine>. Go is planned.

Design principles

  • Two tools, no more. A tiny tool surface keeps the model's context window clean. search_objects progressively discloses schema: call it with no arguments to list tables, with a table name to get columns, indexes, and foreign keys.
  • Safe by default. Read-only mode is enforced in two layers: a conservative SQL guard, plus a session-level read-only setting in the database itself. Rows are capped (default 1000) and statements time out (default 30s).
  • Configured at launch, never via chat. Connection details come from flags, a YAML config file, or environment variables. Credentials are never accepted through a tool call.
  • Secrets never appear in logs. Passwords live in non-printable secret types, DSNs are sanitized before logging, and a redaction filter guards the log boundary.

Quick start

Pick your engine's package; each README has the full config surface. SQLite via npm:

{
  "mcpServers": {
    "sqlite": {
      "command": "npx",
      "args": ["-y", "@database-mcp/sqlite", "--dsn", "/absolute/path/to/database.db"]
    }
  }
}

Or via PyPI: use "command": "uvx" and "args": ["database-mcp-sqlite", "--dsn", "/absolute/path/to/database.db"].

Or as a native binary, with no Node or Python runtime: run cargo install database-mcp-rs-sqlite, then use "command": "database-mcp-sqlite" with the same args minus the package name.

Flags, environment variables, and YAML config are identical across all three lines.

Networked engines take credentials from the environment (MYSQL_*, MARIADB_*, POSTGRES_*/DATABASE_URL, LIBSQL_URL/LIBSQL_AUTH_TOKEN), *_FILE mounted secrets, or a YAML file via --config. Never from a chat prompt.

Contributing

See CONTRIBUTING.md. The short version: the conformance suite is the definition of done. A change is mergeable only when conformance/run.mjs passes against every affected server.

License

MIT