Paritok
A drop-in compression proxy that sits between a coding agent and its LLM API, stripping tool-schema bloat and compressing tool results and stale history to cut input-token bills without changing the agent.
Paritok is a non-destructive compression gateway for coding agents such as Claude Code, Cursor, Codex and OpenHands. An agent points at Paritok instead of the model provider directly; on every request, Paritok strips tool-schema bloat, compresses tool results and file reads, and summarizes stale conversation history before forwarding the request upstream, so the agent is billed on the compressed token count. Nothing is permanently discarded: the agent can pull back the exact original content on demand.
It is powered by an open-source 4B compression model trained specifically on real agent trajectories, and the project reports cutting input-token bills from about 25% on the first turn to more than 85% in long, context-saturated sessions. It works with any agent that honors a configurable BASE_URL, and is open source.