Quickstart

1. Install the package

   pip install llm-parry

   The package is called llm-parry on PyPI, and you import it as parry. It has no required dependencies and hooks whichever SDKs you already have installed.

2. Add one line at startup

   import parry
   parry.install()

   Put it anywhere that runs once, such as main(), an app factory or the top of a notebook. Clients created before and after this line are covered.

3. Make the same call twice

   OpenAI

   import time
   import openai
   import parry
   
   parry.install()
   client = openai.OpenAI()
   
   for i in range(3):
       t = time.perf_counter()
       r = client.chat.completions.create(
           model="gpt-4o-mini",
           messages=[{"role": "user", "content": "Say hi in five words."}],
       )
       print(f"{(time.perf_counter() - t) * 1000:7.1f} ms  {r.choices[0].message.content}")
   
   print(parry.stats().summary())

   Anthropic

   import time
   import anthropic
   import parry
   
   parry.install()
   client = anthropic.Anthropic()
   
   for i in range(3):
       t = time.perf_counter()
       r = client.messages.create(
           model="claude-haiku-5-5",
           max_tokens=100,
           messages=[{"role": "user", "content": "Say hi in five words."}],
       )
       print(f"{(time.perf_counter() - t) * 1000:7.1f} ms  {r.content[0].text}")
   
   print(parry.stats().summary())

   output

     612.4 ms  Hi there, how are you?
       0.4 ms  Hi there, how are you?
       0.1 ms  Hi there, how are you?
   requests        3
   cache           2 hits / 1 misses (67%), 46 tokens and $0.000011 saved
   spend           $0.000006 this session (18 in / 5 out tokens)
   rate limiting   0 delayed, 0 rejected
   retries         0   budget stops 0
     openai:gpt-4o-mini                            1 req  $0.000006

   The first call went to the provider. The next two came from parry’s cache: no network, no cost, and no rate-limit capacity used.

4. Optional: set a budget

   parry.install(max_cost_usd=5.00)

   Once this session has spent $5, every call that isn’t cached raises parry.BudgetExceeded before it leaves your process.

5. Optional: keep everything across restarts

   parry.install(storage="parry.db")

   parry stats --db parry.db

Turning it off

You want	Do
parry off for one block	with parry.bypass(): ...
no caching at all	parry.install(cache=False)
parry in one block only	with parry.Parry() as guard: ... instead of install()
parry on one client only	client = parry.wrap(openai.OpenAI())
undo everything	parry.uninstall()

Identical requests return identical answers

Caching is on by default, so the same request returns the same cached answer even at temperature=1. If you need varied samples, turn caching off for that code path with parry.bypass() or cache=False. See Response caching.