cache-assembler fixes prompt-cache invalidation for Claude Code and the Anthropic API. The cache only hits on byte-identical requests, but real traffic breaks that: inconsistent tool-schema ordering, session data leaking into prompts, concurrent requests racing to write the same entry. This proxy fixes all three, with a dedicated Claude Code preset. Measured against the real API (not synthetic): 8.2x cost reduction, $1.33 to $0.16 per 100-turn session. MIT licensed, zero dependencies.
Fin Startups get Fin free for a year + 93% off Intercom
Promoted
Maker
📌
Hey PH! Built this after realizing my own agentic Claude setup was burning way more tokens than it needed to - turns out prompt caching only works if your request bytes match exactly, and most real usage never hits that bar without help.
cache-assembler sits in front of the API and fixes what normally breaks the match. Measured it properly against the real Anthropic API (not a synthetic benchmark) - 8.2x cost reduction, $1.33 to $0.16 per session.
MIT licensed, zero dependencies. Happy to answer anything about the mechanism or the validation methodology - would love feedback from anyone else running heavy agentic workloads.