llm-cost-optimization

Content-aware output compression for AI coding assistants. 36 specialized processors cut CLI output tokens by 60-99% (git, pytest, npm, terraform, kubectl, docker, and more) without losing errors, diffs, or stack traces.

Open-source, self-hosted AI gateway for multi-tenant orgs: OpenAI-compatible LLM routing, RAG & vector stores, MCP hub, GPU fleet (MIG slicing), sandboxed coding agents, AI red-teaming & guardrails, and cost optimization — all in one console.

Open-source Prompt Firewall — deflect up to 95% of redundant LLM traffic before it leaves your infrastructure. Documents: https://isartor-ai.github.io/Isartor/index.html

Research repository for TokenMizer, a graph-structured session-memory system for long-horizon LLM context management. This repository contains the benchmark suite, the labelled evaluation corpus, all raw results, and the manuscript for the associated paper.

A high-performance, multi-agent observability engine designed for the Model Context Protocol (MCP). It provides a non-blocking, transparent proxy layer that implements deterministic token attribution, real-time context-window alerting, and heuristic-driven static analysis to optimize LLM metadata overhead at scale.

Git-native granular cost accounting for AI-developed repos, what did this commit cost? Reads Claude Code / Codex / Gemini CLI / aider logs, joins them to git history, and prices every commit, PR and release. Measured from real tokens, never estimated.

Right-size every Claude task: advisory slash commands that recommend the cheapest model and effort level whose cost of error is acceptable.

Cost-aware subagent delegation workflow for Cursor: model-pinned dispatch, fail-closed premium gate, and process-driven QA

Open-source, self-hosted Fastest AI security gateway for LLM and agent apps: guardrails, agentic PDP, semantic caching, MCP gateway.

Stop your AI coding agent breaking things. Enforced guardrails that block unsafe edits, composable modes, and token-cost budgeting for Kiro CLI.

Building production-grade AI infrastructure — RAG, agents, evals, and cost-aware LLM systems. Currently: routing queries to the cheapest model that can handle them.

3,300 practical techniques to cut LLM token usage by up to 90% — 300 core principles + 3,000 context-specific applications.

Local-first caching proxy for OpenAI-compatible LLM APIs: exact, semantic and tool-result caching plus request coalescing to cut agent costs

Builds prompt-cost discipline into Claude Code: measure every prompt, surface avoidable spend, report org-wide. Skill + UserPromptSubmit hook + MCP rollup with cited dollar/Wh/CO2e receipts.

Sample code for two-tier LLM inference (cascading) on Amazon Bedrock, using a single Bedrock Converse API surface. NVIDIA Nemotron Nano handles routine support-ticket classification on every request; Anthropic Claude Sonnet handles cases the routing logic flags as harder. Includes a bake-off harness, sample data, and tests. Next.js + TypeScript.

Tamper-evident, stranger-verifiable receipts for LLM cost-savings — anyone can recompute your caching/routing savings math offline, no trust in your dashboard required. Pure stdlib, zero-dependency.

AI cost governance toolkit for Claude Code. Monitor token usage, prevent runaway AI agent costs, optimize prompts, and enforce intelligent spending limits for enterprise AI workflows.

Cut Claude Code subagent unit cost 3x in 26 measured days: one policy table, two pinned agents, a 15-line deny hook. No classifier, no proxy — receipts included.

A lightweight OpenClaw model router that reduces LLM costs by switching requests to cheaper models using custom rules.