#75 · Primary category: AI Gateway & API Infrastructure

paritok-4b-v1

agentic-coding ai ai-agents ai-coding ai-infrastructure developer-tools llm-gateway llms token-optimization

Non-destructive compression gateway for AI coding agents, cutting token bills up to 85% and fitting ~3× more turns in the same context window.

Project last updated:08/29/26

GitHub Stars

1.4K

Forks

137

Contributors

7

License

Apache-2.0

Why we included this project

Long agentic coding sessions burn input tokens on the same tool schemas and file reads every turn, and the bill climbs as the context window fills. Paritok sits between your agent and the upstream model as a drop-in proxy that honors BASE_URL, so Claude Code, Cursor, Codex, OpenHands, or any agent that points at a custom endpoint can use it with zero changes to the agent itself. Each request is filtered for irrelevant tool schemas, and the tool results and file reads are compressed by a code-native 4B model trained on real coding trajectories, with stale history summarized once the window fills; nothing is permanently discarded, so the agent can pull back any exact original on demand. Savings start around 25% on the first turn and pass 85% in long, context-saturated sessions, which roughly triples the number of turns you fit in the same window. It self-hosts, with the embedding filter running locally on CPU, so teams already paying for agentic coding tools get a concrete way to stretch both budget and context without reworking their setup.

Articles for this project

No articles for this project yet.

To suggest a topic or contribute an article, contact us.

Related projects in this category