Paritok 4b V1

This is a compression gateway for AI coding agents. It sits between a coding agent and the language model provider and rewrites requests on the fly, compressing tool schemas, file reads, command output and conversation history so that sessions run longer and token bills are smaller. Integration is a single environment variable that points the agent's base URL at the gateway, and the vendor lists support for Claude Code, Cursor, Codex, OpenHands and any OpenAI compatible upstream. The approach has three parts. Tool schema filtering keeps the tools relevant to a turn and stubs the rest, which the vendor says reduces a typical 29K token tool block to about 8K. File reads and tool output are compressed by the vendor's open source, code native 4 billion parameter model, which is designed to keep identifiers, paths and errors intact. Older history is summarized as the context window fills. Compression is described as non destructive: touched content is tagged, and a read_original reference returns the exact original bytes on demand. Responses come back untouched. The site includes a savings calculator by model and context size. Pricing is usage based. The gateway can be self hosted for free. A hosted GPU option costs $0.30 per 1 million billable file tokens using prepaid credit, with $5 of free credit for new accounts and no card needed, and tool schema filtering is always free. Credit can be topped up from $5.