# Paritok-official/paritok-4b-v1

Non-destructive compression gateway for AI coding agents. Cuts token bills 25% on turn 1 to past 85% in long or saturated sessions, and fits ~3× more turns in the same context window. Powered by our open-source code-native 4B model. Drop-in for Claude Code, Cursor, Codex, OpenHands, and any BASE_URL agent.

Repository: https://github.com/Paritok-official/paritok-4b-v1
Canonical: https://ross.abutalabs.com/products/paritok-4b-v1
Homepage: https://www.paritok.com/
Language: Python
License: Apache-2.0
License Family: permissive
Topics: ai, ai-infrastructure, developer-tools, llms, agentic-coding, ai-agents, ai-coding, llm-gateway, token-optimization
Last push: 2026-08-24T11:57:58+00:00

## Health v2 (maintenance only)
Score: 57/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 99, release rhythm 35, longevity 3
- inputs: {"age_days": 49, "days_push": 9, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases, young
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1449, forks 138 (observed 2026-08-28T04:04:45.725275+00:00)

## What it is
Paritok is a non-destructive compression gateway that sits as a drop-in proxy between AI coding agents (Claude Code, Cursor, Codex, OpenHands, or any BASE_URL-configured agent) and their LLM upstream. It uses an open-source 4B model (Qwen3-4B backbone) to strip tool-schema bloat, compress tool results and file reads, and summarize stale history, cutting input-token bills ~25% on turn one to past 85% in long sessions while allowing exact originals to be pulled back on demand.

## Use cases
- reduce LLM API token costs for AI coding agents
- fit more turns into a fixed context window
- compress tool schemas and file reads in agent sessions
- self-host a proxy to cut Claude Code or Cursor token bills
- summarize stale conversation history without losing recoverable originals
- recover exact original file content compressed by the gateway
- filter irrelevant tool definitions per request to shrink prompts

## When to choose
- you run coding agents like Claude Code, Cursor, Codex, or OpenHands and want lower token bills without changing agent code
- your agent sessions hit context-window saturation and you need ~3x more turns in the same window
- you want a self-hosted, Apache-2.0 compression layer with on-demand recovery of original content

## When to avoid
- you need zero added latency or cannot run a local 4B model / proxy hop
- your workflow requires byte-exact prompts upstream at all times with no rewriting
- you only make short, single-turn LLM calls where compression savings are minimal

## Facets
- artifact type: service
- maturity: active
- function: llm-inference, proxy, middleware, machine-learning, developer-tools
- domain: large-language-models, developer-tools, infrastructure-as-code, self-hosted
- platform: python, self-hosted, cross-platform
- tags: llm-gateway, token-optimization, context-compression, coding-agents, drop-in-proxy, cost-reduction, qwen3-4b, prompt-compression, ai-agents, docker

## Member repositories
- Paritok-official/paritok-4b-v1 (main) score 57

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:04:45.725275+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T04:35:58.959294+00:00, confidence not recorded.
  - readme: https://github.com/Paritok-official/paritok-4b-v1 (fetched 2026-08-28T04:04:45.725275+00:00, sha 1798284054a5)
  - homepage: https://www.paritok.com/ (fetched 2026-08-29T11:45:35.463957+00:00, sha 87382babe42b)
- Data as of 2026-08-30T08:39:29.467469+00:00.