Configuration

Every option has a default. You only need this page when the defaults don’t fit.

`Parry(...)`

from parry import Parry

guard = Parry(
    cache=True,
    cache_ttl=3600,
    cache_size=10_000,
    rate_limit=True,
    on_rate_limit="wait",
    max_wait=60.0,
    cost_tracking=True,
    max_cost_usd=None,
    retry=True,
    max_attempts=3,
    retry_base_delay=0.5,
    retry_max_delay=30.0,
    storage=None,
)

Option	Type	Default	Meaning
`cache`	`bool`	`True`	serve identical requests from the cache
`cache_ttl`	`float \| None`	`3600`	seconds an entry lives. `None` means it lives until evicted by size
`cache_size`	`int`	`10_000`	maximum number of entries. The cache is always bounded, so `None` or `0` raises `ValueError`
`rate_limit`	`bool`	`True`	apply per-model token buckets
`on_rate_limit`	`"wait" \| "raise"`	`"wait"`	sleep until a slot frees, or raise `RateLimited` immediately
`max_wait`	`float`	`60.0`	in `"wait"` mode, raise `RateLimited` once waiting would exceed this many seconds
`cost_tracking`	`bool`	`True`	record spend in the ledger
`max_cost_usd`	`float \| None`	`None`	hard budget for this guard’s session. See Budgets
`retry`	`bool`	`True`	retry transient failures
`max_attempts`	`int`	`3`	total attempts, including the first
`retry_base_delay`	`float`	`0.5`	first backoff delay in seconds. It doubles on each attempt, with equal jitter
`retry_max_delay`	`float`	`30.0`	cap for any single delay, including server `Retry-After` values
`storage`	`str \| Path \| None`	`None`	SQLite file for a persistent cache and ledger

cache, rate_limit, cost_tracking, retry, max_cost_usd, on_rate_limit and max_wait are plain attributes, so you can change them at runtime:

guard.cache = False
guard.max_cost_usd = 25.0

`parry.install(...)`

parry.install(guard=None, *, sync=True, async_=True, openai=True, anthropic=True, **options)

Argument	Default	Meaning
`guard`	`None`	install this `Parry` instead of creating one
`sync`	`True`	hook sync clients
`async_`	`True`	hook async clients (`parry.install_async()` hooks async only)
`openai`	`True`	hook the OpenAI SDK
`anthropic`	`True`	hook the Anthropic SDK
`**options`		passed to `Parry(...)` when a new guard is created

Per-model settings

guard.set_rate_limit(model, requests_per_minute=None, tokens_per_minute=None, *, burst=None, provider=None)
guard.remove_rate_limit(model, *, provider=None)
guard.set_pricing(model, input_per_million, output_per_million)

See Rate limiting and Cost tracking.

Cache backend

guard.set_cache("memory")                    # default
guard.set_cache("sqlite", path="parry.db")
guard.storage("parry.db")                    # same as the line above

Environment

parry reads no environment variables and no config files. Everything is set in code. Your SDKs keep reading their own variables as usual (OPENAI_API_KEY, ANTHROPIC_BASE_URL, …).