#token-optimization

Cut context bloat in your AI-agent stack: find and safely prune unused skills, MCP servers and subagents from real transcript evidence

MCP-native code retrieval for AI agents — 84-88% fewer read tokens, BM25F + semantic search, AST chunks, session dedup

Agentic Runtime

Gobstopper makes long Claude Code and Codex sessions smaller. A local proxy compacts live requests, and saved sessions get a smaller copy beside the original.

Self-hosted agentic OS — a Vibe-Engineering platform for Builders. Connect Claude Code, Codex or Hermes Agent to Discord, Telegram, WhatsApp, Slack & Email. EU AI Act 2026 & GDPR compliance by architecture.

Windows fork of pxpipe: transparent Fable 5 compression for Claude Desktop on Windows 11.

Safe, quality-gated model+effort router for Claude Code - delegates cheap work down to save tokens, escalates hard/sensitive work up through a verify staircase, never-down gate, and fail-safe Lock, up to a consent-gated fable top rung. Not the cheapest router; the safest one inside Claude.

Enhanced fork of OpenWolf — a token-conscious second brain for Claude Code, with bounded storage, self-maintenance (openwolf doctor), .wolfignore scoping, and tunable retention. AGPL-3.0.

Local orchestration for Claude Code — session recovery, scheduling, worktrees, Token Saver, and dashboards via official CLI transport. Independent third-party software.

Local-first agentic memory lifecycle manager for AI coding IDEs

Local-first Build Truth & verification layer for AI coding agents. Stops hallucinations, cuts token costs by 90%, and fixes monorepo test failures for Claude, Cursor, Aider, & Copilot.

Auto-route each Claude Code session to Haiku, Sonnet, Opus or Fable from its first prompt. Classified locally, zero tokens.

The non-obvious setup that keeps a Claude Code project sharp as it grows: skill architecture, context budget, document ingestion, and permission/trust. The things I wish I knew before I started.

Token optimizer for AI coding agents - Claude Code, Codex, Cursor, Aider, Gemini CLI, Cline, Windsurf, Copilot.

Cut LLM token consumption 65-80% in Hermes Agent — RTK (tool-output) + Caveman (agent-output) compression plugins, one-command install.

🪨 why input many code when few code do trick. Slashes input tokens by up to ~80% across 9 languages via AST pruning, ultra-dense "caveman" syntax and more for AI coding agents.

When Jev meets LLM — Transparent proxy that makes AI coding agents 3× faster and 80% cheaper. Cut response times from 2.4s to 890ms with Line J architecture.

Cut Claude Code token costs ~80% by routing analysis to DeepSeek and graphify before reading source files.

PECS: 基于 LangGraph 的四角色多智能体任务求解框架（Planner/Executor/Critic/Synthesizer）。WebShop 真实环境 +25pp (25% vs 0%)；GAIA 官方 53 题 26.4% vs ReAct 24.5%（McNemar 不显著）。含 AST 沙箱、50000 token 硬预算、FastAPI 限流/混沌/CI/Prometheus。

Offload coding-agent busywork to a local llama.cpp server; zero cloud tokens. Toolkit + Qoder/copilot/claude CLI skill/agent integration.

Local-first context cleaner for AI agents.

Sub-second local business telemetry & flaw audits in ~210 tokens. Official Model Context Protocol (MCP) server & CLI for Claude Desktop, Cursor, and AI agents.

Sub-150ms Web-to-AST Markdown engine & Model Context Protocol (MCP) server. Eliminates 90%+ raw HTML token bloat for LLMs, RAG, and agent context.

In-flight conversational context compressor and stale-turn pruner to eliminate context rot.

Obsidian-based persistent memory framework for AI coding agents. Set it up with just a single prompt.

Plug-and-play Copilot agent bundle that cuts token usage ~50% via orchestrator-worker context isolation

A personal baseline template for Claude Code: spend fewer tokens per session without touching answer quality. Built for Sonnet, high effort.

Experimental opt-in Jev skills and config for Claude Code — the companion to jev-codex-tools.

Set each Claude Code sub-agent's reasoning effort per spawn: start the Agent prompt with [effort: low|medium|high|xhigh|max].

engramx's bi-temporal mistakes + structural code-graph as auto-discoverable Claude Code Skills. Catches recurring mistakes before agents repeat them. Apache 2.0, local SQLite, zero cloud.

A Claude skill that cuts token waste lean replies, smart file reading, zero redundant work

Local-first context gateway that helps AI coding agents read less, spend less, and stay grounded in your repo.

Reduce Codex and Claude context waste by measuring MCP tools, repo instructions, and safe token-budget cuts.

Deterministic Claude Code hook system that reduces real token consumption (no extra LLM calls) via read dedup, diff-serving, Bash output distillation, large-file outlines, compaction awareness, and capped cross-session memory.

Local context-compression layer for agent tool outputs. Claude Code PostToolUse hook or generic filter, single Go binary.

Measure Claude Code token use, cut the waste, and prove your savings with a real before and after. Local, private, measured from your own usage counters.

Claude Code hooks that measure live context, warn as it grows, and force a detailed handoff note before a long chat gets expensive.

Confidence-gated model routing for coding agents — run work at the cheapest model that can pass verification. SKILL.md, MIT.