mcp-sqlite-tools

A Model Context Protocol (MCP) server for safe, local SQLite database operations. It gives MCP clients explicit read, write, schema, transaction, CSV, backup, and maintenance tools.

Features

• Open, create, inspect, back up, vacuum, and close SQLite databases
• List, describe, create, and drop tables
• Run paginated read queries with named or positional parameters
• Run explicit write and schema queries
• Import and export headered CSV files
• Use transactions with nested savepoints
• Export and import schemas as SQL or JSON
• Restrict database and CSV paths through configuration
• Classify destructive tools for client approval policies
• Use Node's built-in SQLite driver with no native addon dependency

Requirements

• Node.js 24.12 or later
• An MCP client with stdio server support

Configure your MCP client

The MCP client starts this server through npx; a global installation is not required.

Install with MCPick

MCPick can add the server to a supported client without manual JSON editing. This command targets Claude Code's local scope by default:

npx -y mcpick add \
  --name sqlite-tools \
  --command npx \
  --args "-y,mcp-sqlite-tools"

Select a client and scope explicitly when needed:

npx -y mcpick add \
  --name sqlite-tools \
  --command npx \
  --args "-y,mcp-sqlite-tools" \
  --client vscode \
  --scope project

The add command supports Claude Code, Gemini CLI, VS Code, Cursor, Windsurf, OpenCode, and Pi. Run npx mcpick clients to see current client capabilities, scopes, and configuration locations.

The examples track the latest package release. For reproducible configuration, replace mcp-sqlite-tools in --args with an exact version such as [email protected].

Manual configuration

For unsupported clients or advanced configuration, add the server manually:

{
	"mcpServers": {
		"mcp-sqlite-tools": {
			"command": "npx",
			"args": ["-y", "mcp-sqlite-tools"],
			"env": {
				"SQLITE_DEFAULT_PATH": ".",
				"SQLITE_ALLOW_ABSOLUTE_PATHS": "true",
				"SQLITE_BUSY_TIMEOUT": "30000",
				"SQLITE_JOURNAL_MODE": "wal",
				"SQLITE_BACKUP_PATH": "./backups"
			}
		}
	}
}

VS Code uses a servers object instead of mcpServers. See the configuration guide for more client-specific examples.

Environment variables

| Variable | Purpose | Default | | ----------------------------- | ----------------------------------------- | --------------------- | | SQLITE_DEFAULT_PATH | Base directory for database files | . | | SQLITE_ALLOW_ABSOLUTE_PATHS | Allow paths outside the default directory | true | | SQLITE_BACKUP_PATH | Default backup directory | Default database path | | SQLITE_BUSY_TIMEOUT | SQLite lock busy timeout in milliseconds | 30000 | | SQLITE_JOURNAL_MODE | delete, truncate, persist, or wal | wal | | DEBUG | Enable diagnostic logging | false |

When SQLITE_ALLOW_ABSOLUTE_PATHS=false, relative and absolute paths must resolve inside SQLITE_DEFAULT_PATH. Symbolic links cannot be used to escape that directory.

Use SQLITE_JOURNAL_MODE=delete for databases in file-synchronized directories. The unsafe SQLite modes memory and off are not accepted.

SQLITE_MAX_QUERY_TIME remains available as a deprecated alias for SQLITE_BUSY_TIMEOUT. It does not limit wall-clock query runtime.

Tools

Tools are separated by intent so MCP clients can apply clear approval rules.

Safe and read-only

| Tool | Purpose | | -------------------- | ---------------------------------- | | open_database | Open an existing database | | close_database | Close one database connection | | list_databases | Find database files in a directory | | database_info | Read file and SQLite metadata | | list_tables | List tables and views | | describe_table | Read columns and constraints | | backup_database | Create a consistent online backup | | export_csv | Export a table or read-only query | | export_schema | Export schema as SQL or JSON | | execute_read_query | Run one SQLite read-only statement |

Writes, schema, and maintenance

| Tool | Purpose | | ---------------------- | --------------------------------------- | | create_database | Create a new database file | | create_table | Create a table from validated columns | | drop_table | Drop a table and its data | | execute_write_query | Run INSERT, UPDATE, or DELETE | | execute_schema_query | Run one schema statement | | bulk_insert | Insert records in batches | | import_csv | Import a headered CSV file | | import_schema | Apply SQL or JSON schema objects | | vacuum_database | Rebuild a database to reclaim space | | begin_transaction | Begin a transaction or nested savepoint | | commit_transaction | Commit or release a savepoint | | rollback_transaction | Roll back a transaction or savepoint |

See the complete API reference for parameters, responses, examples, pagination, and CSV options.

Safety model

The server does not treat every SQL string as equivalent:

• execute_read_query uses SQLite's authorizer API to reject writes, schema changes, unsafe PRAGMAs, attachment, and multiple statements.
• Write, schema, transaction, and destructive administration actions use separate tools so clients can request approval.
• Database and CSV paths are resolved and validated before access.
• Identifiers generated by tools are quoted.
• Values are bound as parameters rather than interpolated into SQL.
• Backups use SQLite's online backup API and include committed WAL data.

A client can allow read-only tools and require approval for destructive tools. Always review SQL and file paths before approving changes. Back up important databases before schema changes, imports, or large writes.

Why native SQLite?

Using node:sqlite removes the native addon, its install script, and its platform-specific binaries. A clean production install fell from 31.3 MB with better-sqlite3 to 3.6 MB with native SQLite, an 88.5% reduction. The npm tarball itself is similar in size: 63.1 KB native versus 60.0 KB published. The large saving is in the installed dependency tree.

The migration also removes better-sqlite3 and its type package. It makes installation independent of prebuilt addon availability or a working native compiler.

Driver benchmark

Lower times are better. These medians use 20,000 rows, two warmups, and seven measured runs per driver. Each sample uses a new database and the driver order alternates. Setup is outside the measured region except for the insert workload.

| Workload | node:sqlite | better-sqlite3 | Native result | | ------------------- | ------------: | ---------------: | ------------: | | Insert transaction | 11.13 ms | 23.78 ms | 2.14× faster | | Indexed point reads | 25.23 ms | 21.40 ms | 17.9% slower | | Full row scan | 6.39 ms | 3.29 ms | 94.3% slower | | Update transaction | 8.04 ms | 15.40 ms | 1.92× faster | | Online backup | 0.55 ms | 0.38 ms | 45.0% slower |

Measured on Linux x64 with Node.js 24.15.0 and an AMD Ryzen AI 9 HX 370. Node used SQLite 3.51.3; [email protected] used SQLite 3.53.3. These microbenchmarks show driver trade-offs, not complete MCP performance. MCP transport and validation costs are not included.

Development

git clone https://github.com/spences10/mcp-sqlite-tools.git
cd mcp-sqlite-tools
pnpm install
pnpm run check
pnpm test
pnpm run build

See development and architecture for module responsibilities and other development commands.

Documentation

• Configuration
• API reference
• Development and architecture

Contributing

Issues and pull requests are welcome.

License

MIT License. See LICENSE.