run-sql-connectorx

An MCP server that executes SQL via ConnectorX and streams the result to CSV or Parquet in PyArrow RecordBatch chunks.

• Output formats: csv or parquet
• CSV: UTF-8, header row is always written
• Parquet: PyArrow defaults; schema mismatch across batches raises an error
• Return value: the string "OK" on success, or "Error: " on failure
• On failure the partially written output file is deleted
• CSV token counting (optional): per-line token counting via tiktoken (o200k_base) with a warning threshold

Why this library?

• Efficient streaming: handles large results in Arrow RecordBatch chunks
• Token-efficient for MCP: exchanges data via files instead of inline payloads
• Cross-database via ConnectorX: one tool works across many backends
• Robust I/O: CSV header handling, Parquet schema validation, safe cleanup on errors

Supported data sources (ConnectorX)

ConnectorX supports many databases. Common examples include:

• PostgreSQL
• MySQL / MariaDB
• SQLite
• Microsoft SQL Server
• Amazon Redshift
• Google BigQuery

For the complete and up-to-date list of supported databases and connection-token (conn) formats, see the official docs:

• ConnectorX repository:
• Database connection tokens:

Getting Started

uvx run-sql-connectorx \
  --conn "" \
  --csv-token-threshold 500000

is the connection token (conn) used by ConnectorX—SQLite, PostgreSQL, BigQuery, and more.

CLI options

• --conn (required): ConnectorX connection token (conn)
• --csv-token-threshold (default 0): when > 0, enable CSV per-line token counting using tiktoken(o200k_base); the value is a warning threshold

Further reading

• ConnectorX repository:
• Connection-token formats for each database:

Running from mcp.json

To launch the server from an MCP-aware client such as Cursor, add the following snippet to .cursor/mcp.json at the project root:

{
  "mcpServers": {
    "run-sql-connectorx": {
      "command": "uvx",
      "args": [
        "--from", "git+https://github.com/gigamori/mcp-run-sql-connectorx",
        "run-sql-connectorx",
        "--conn", ""
      ]
    }
  }
}

Behaviour and Limits

• Streaming: Results are streamed from ConnectorX in RecordBatch chunks; the default batch_size is 100 000 rows.
• Empty result:
  • CSV – an empty file is created
  • Parquet – an empty table is written
• Error handling: the output file is removed on any exception.
• CSV token counting (when --csv-token-threshold > 0):
  • Counted text: exactly what csv.writer writes (including header row when present, delimiters, quotes, and newlines), UTF-8
  • Streaming approach: tokenized with tiktoken(o200k_base) per written CSV line

Call output

The tool returns a single text message.

• On success:
  • Parquet: OK
  • CSV:
    • If --csv-token-threshold = 0: OK
    • If --csv-token-threshold > 0: OK N tokens (or OK N tokens. Too many tokens may impair processing. Handle appropriately when N >= threshold)
    • Empty result with counting enabled: OK 0 tokens
• On failure: Error: (any partial output file is deleted)

MCP Tool Specification

The server exposes a single MCP tool run_sql.

Argument	Type	Required	Description
sql_file	string	yes	Path to a file that contains the SQL text to execute
output_path	string	yes	Destination file for the query result
output_format	enum	yes	One of "csv" or "parquet"
batch_size	int	no	RecordBatch size (default 100000)

Example Call

{
  "tool": "run_sql",
  "arguments": {
    "sql_file": "sql/queries/sales.sql",
    "output_path": "output/sales.parquet",
    "output_format": "parquet",
    "batch_size": 200000
  }
}

---

License

Distributed under the MIT License. See LICENSE for details.