API reference

Generated from the source, so it cannot drift from the code. Every signature here is the real one — the package is checked with mypy --strict and ships a py.typed marker, so your own type checker sees these types too.

The surface a caller touches

SQL	connect, the connection, .query(...), .tool(), and SQLPolicy
Results and outputs	what an operation hands back: inline, outputs, uri, status, metrics
Failures	the normalized failure taxonomy, and which kinds are retryable
Handles and execution	reconnecting to work in flight, polling it, cancelling it
Batch and stream jobs	gantry.batch and gantry.stream
Capability matrix	which policy fields each provider can actually enforce

The shape of every operation

Configuration and enforcement are separated on purpose. Application code holds the credential and decides policy and trusted checks; the agent receives a tool whose inputs are the SQL and an optional allowlisted verification commitment.

import gantry

db = gantry.sql.connect("postgres", url=...)  # application code
query = db.query(read_only=True, schemas=("analytics",))  # application code
tool = query.tool()  # what the agent gets

tool.input_schema contains sql and verify. The latter exposes only declarative checks supported by that operation/provider. Policy and trusted checks are not arguments a model can pass, so the model cannot widen authority or weaken the application acceptance contract.