Paritok

Paritok is a self-hosted gateway for developers whose coding agents send large file reads, tool output and conversation histories to an LLM. It reduces that input before the request reaches the upstream model.

• Smaller agent context: It trims tool definitions, compresses file reads and tool results, and summarizes older conversation history.
• Original text on demand: Agents can retrieve the exact text behind a compressed passage when they need details the shorter version left out.
• Coding-agent compatibility: It works with Claude Code, Cursor, Codex, OpenHands and agents that support a custom base URL.
• Code-focused compression: Its Paritok-4B-v1 model selects passages rather than rewriting them and is trained to retain details such as function names, paths and error strings.

The gateway and model are free and open source under Apache 2.0. Local compression runs through Ollama or vLLM; vLLM can serve the Qwen3-4B-Instruct-2507-based adapter on a 24GB GPU. Self-hosting needs no Paritok key and has no telemetry. Compressed requests still go to the chosen upstream LLM, such as Anthropic or an OpenAI-compatible provider. The paid hosted option sends content to Paritok's GPU server for compression and requires an API key.