paritok-official/paritok-4b-v1 — explained in plain English
Analysis updated 2026-08-13
Cut AI API token costs for coding agents like Claude Code, Cursor, or Codex.
Fit more turns into a long agent session without hitting the context window limit.
Reduce repeated tool-schema overhead sent on every agent request.
Recover the exact original file or tool output when the agent needs full detail.
| paritok-official/paritok-4b-v1 | openmoss/moss-transcribe-diarize | jordan-gibbs/hyperresearch | |
|---|---|---|---|
| Stars | 1,142 | 1,139 | 1,133 |
| Language | Python | Python | Python |
| Setup difficulty | moderate | hard | easy |
| Complexity | 3/5 | 4/5 | 3/5 |
| Audience | developer | developer | researcher |
Figures from each repo's GitHub metadata at analysis time.
Point your existing agent's BASE_URL at the Paritok gateway, no code changes needed.
Paritok is a middle layer that sits between an AI coding agent, such as Claude Code, Cursor, or Codex, and the AI model it talks to. Instead of changing anything about how you use your coding agent, you point it at Paritok instead of pointing it directly at the AI provider, and Paritok quietly shrinks the amount of text sent on every request before passing it along. The tool works through three separate techniques. First, it filters the list of tools an agent has access to. Coding agents often expose dozens of available tools in full detail on every single request, even though only a few are relevant to the current task, so Paritok keeps just the relevant ones in full detail and shrinks the rest, which can be recovered later if actually needed. Second, it compresses file contents and tool results, shrinking them to roughly a quarter of their original size while trying to preserve important details like function names, file paths, and error messages. Third, once a conversation grows long enough to fill up the model's available context, it summarizes older parts of the conversation so the session can keep going instead of running out of room. Importantly, none of this deletes information permanently. If the agent needs the exact original text it filtered or shrank, it can request it back on demand. The project reports that this approach can cut the amount of billed input text by roughly a quarter on the very first request, and by well over 80 percent in long running sessions where the same information would otherwise be resent over and over. The underlying compression is powered by a small, openly released 4 billion parameter AI model trained specifically for this task, built on the Qwen3-4B model and available on Hugging Face. The project is released under the Apache 2.0 license, which allows free use, modification, and commercial use. This tool is aimed at developers who use AI coding agents regularly and want to reduce their API costs and fit more conversation turns into a single session without switching tools.
A proxy that sits between AI coding agents and the LLM API, compressing requests to cut token costs by 25 to over 85 percent without losing information.
Mainly Python. The stack also includes Python, Qwen3-4B.
Use, modify, and distribute freely, including for commercial purposes, with patent protections and a requirement to keep license notices.
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly developer.
This repo across BitVibe Labs
Don't trust strangers blindly. Verify against the repo.