Changelog

0.1.0 2026-10-25

The first release. See also CHANGELOG.md and GitHub releases.

Hooks

• parry.install() hooks OpenAI chat completions, the OpenAI Responses API, Anthropic messages and Anthropic beta messages, on sync and async clients, including .with_raw_response and stream=True.

Guard

• Response cache: a blake3 key over the canonicalised semantic request, a bounded moka LRU with TTL, and optional SQLite persistence.
• Per-model token-bucket rate limiting with tier-1 defaults, "wait" or "raise" semantics, burst and tokens-per-minute limits.
• Cost tracking from a dated pricing table, matched by longest prefix, with per-model and per-day totals.
• max_cost_usd budget, checked before each request is sent.
• Retry with exponential backoff, equal jitter and Retry-After.

Scoping

• with Parry():, parry.wrap(client), parry.bypass().

Integrations

• LangChain ParryCallback and LlamaIndex ParryHandler.

Tooling

• The parry stats and parry pricing CLI.
• Standalone Rust primitives: Cache, RateLimit, Cost, Retry, count_tokens and estimate_tokens.
• Published on PyPI as llm-parry and imported as parry.