strands-sql

A general-purpose SQL tool for Strands Agents — supports PostgreSQL, MySQL, and SQLite via SQLAlchemy.

Installation

# SQLite (no extra driver needed)
pip install strands-sql

# PostgreSQL
pip install "strands-sql[postgres]"

# MySQL
pip install "strands-sql[mysql]"

strands-sql requires sqlglot for SQL parsing — it is installed automatically as a dependency.

Quick Start

from strands_sql import StrandsSQL

db = StrandsSQL("sqlite:///./local.db")

print(db.list_tables())
print(db.schema_summary())
print(db.describe_table("users"))
print(db.query("SELECT * FROM orders WHERE amount > 100"))

# Write data (disabled by default — pass read_only=False to enable)
db_write = StrandsSQL("sqlite:///./local.db", read_only=False)
db_write.execute("INSERT INTO users (name, age) VALUES ('Eve', 22)")

Use with a Strands Agent

from strands import Agent
from strands_sql import StrandsSQL

db = StrandsSQL("sqlite:///./local.db")

# Use db.as_tool() to preserve your connection and settings
agent = Agent(tools=[db.as_tool()])

agent("How many users are there?")
agent("Show me all orders above 100")
agent("What tables exist in this database?")

⚠️ Note
Always use db.as_tool() rather than passing sql_database directly. as_tool() binds your connection string, read_only flag, table access rules, and other settings to the tool — passing sql_database directly means the agent must supply all of these itself on every call.

Configuration

Connection String

Pass it to StrandsSQL() directly, or set the DATABASE_URL environment variable:

export DATABASE_URL="postgresql://user:password@localhost:5432/mydb"

db = StrandsSQL("postgresql://user:password@localhost:5432/mydb")  # explicit
db = StrandsSQL()  # reads DATABASE_URL automatically

Options

db = StrandsSQL(
    "sqlite:///./local.db",
    read_only=True,
    max_rows=500,
    timeout=30,
    output_format="markdown",
    allowed_tables=["users", "orders"],
    blocked_tables=["secrets"],
)

Methods

list_tables()

List all accessible tables and views.

describe_table(table)

Show columns, types, primary keys, and foreign keys for a table.

schema_summary()

Compact schema of all tables — ideal for giving an LLM context about your database.

query(sql, *, output_format=None, max_rows=None)

Run a SELECT statement. Both output_format and max_rows can be overridden per-call. Write queries are blocked when read_only=True.

db.query("SELECT * FROM users")                          # markdown (default)
db.query("SELECT * FROM users", output_format="json")   # JSON array
db.query("SELECT * FROM logs", max_rows=100)            # override row cap

execute(sql)

Run a write statement (INSERT / UPDATE / DELETE / DDL).

Raises PermissionError if read_only=True. If allowed_tables or blocked_tables are configured, access rules are still enforced and return an error string rather than raising.

db_write = StrandsSQL("sqlite:///./local.db", read_only=False)
db_write.execute("INSERT INTO users (name, age) VALUES ('Eve', 22)")
db_write.execute("UPDATE users SET age = 30 WHERE name = 'Alice'")
db_write.execute("DELETE FROM users WHERE name = 'Bob'")

as_tool()

Return a Strands-compatible tool bound to this instance's settings.

Output Formats

db.query("SELECT * FROM users", output_format="markdown")  # default
db.query("SELECT * FROM users", output_format="json")

Low-level API

For advanced use cases, two additional functions are available:

• get_tool() — returns a Strands Tool that reads DATABASE_URL from the environment at call time. Useful when you don't want to construct a StrandsSQL instance.
• run_sql_database(**kwargs) — calls the tool handler directly without the ToolUse wrapper format. Prefer StrandsSQL for new code.

Development

git clone https://github.com/NithiN-1808/strands-sql
cd strands-sql
pip install -e ".[dev]"
pytest
pytest --cov=strands_sql --cov-report=term-missing
ruff check src/ tests/
mypy src/strands_sql/

License

Apache 2.0