llm-cost-reduction 开源项目推荐（16 个）

open-compress /claw-compactor

14-用于LLM令牌压缩的Fusion Pipeline阶段-可逆压缩、AST感知代码分析、智能内容路由。零LLM推理成本。麻省理工学院获得许可。

HoangP8 /tokless

一个统一的工具包，用于高效且有效的编码代理（Karpathy原则、Caveman、Ponytail、TEK、CodeCurve、Context-Mode）。最短设置时间不超过30秒。任何操作系统。

arjunkshah /supercompress

Neural context compression for long-running AI agents. Query-aware context compiler.

smartass-4ever /Mnemon

Biological nervous systems don't recompute known workflows from scratch. Mnemon gives LLM agents the same primitive — execution memory that caches plans, not responses. 93% token reduction, 2.66ms vs 20s, zero tokens on repeat runs. LangChain, CrewAI, AutoGen.

chaitanya-sharmaa /axon

该项目在 GitHub 上没有填写描述

ramaikn /ponycave

The only token-saving agent rules you may need.

SynStar-Joey /llm-cost-calculator

Simple tool to estimate AI API costs when using large language models.

AAH20 /apex-token-slasher

Deterministic Context Compression & Submodular Call-Graph Pruner — Cuts LLM Bills by 75-85% in Pure Python

marceloasl99 /canonical-rag

Ultra-lightweight (~40MB RAM), Zero-GPU Two-Stage RAG engine powered by SQLite FTS5, local FlashRank cross-encoder, and semantic caching.

CaiqueDourado /llm-cost-kit

What each LLM call really cost: prices with source and date, thinking tokens, price tiers, batch and off-peak, a ledger that never breaks your pipeline

martianzhang /json2toon

Convert JSON format to TOON

Aravind7262 /content-pipeline

Automate content research, card news, images, voice, and video from one prompt with an end-to-end Claude Code content pipeline

shadialhasan /eidcloud-arabic-token-optimizer

High-efficiency Arabic text compressor and token optimizer reducing LLM token billing by 25-40% in pure PHP

ishandutta2007 /Awesome-Claude-Bill-Optimizer

该项目在 GitHub 上没有填写描述

lavs-ai /agent-budget

Experimental CLI for estimating AI-agent request costs and tracking budget assumptions across OpenAI, Anthropic and GitHub Copilot.

blkdynamite /Gio-prompt-optimizer

Free Claude Code plugin: targeted file:line retrieval instead of whole-file reads. Measured 6–14× fewer input tokens.