MCP servers that ship as code — a repository or a package you install and run locally (usually over stdio). They have no hosted endpoint, so we cannot probe or grade them: instead each one shows its repository, install command, license, stars and when it was last updated.
Collected from GitHub repositories and npm, PyPI, Cargo, Go, Maven, NuGet and RubyGems packages tagged as MCP servers, the official MCP Registry, Docker's MCP catalog, ModelScope, Baidu MCP World, Cline and other public directories; one entry per repository. A server that also offers a hosted endpoint is listed under Remote instead. These entries are not counted in our reachability, verification or readiness statistics.
-
@iflow-mcp/flightlesstux-prompt-caching-mcpgithub.com/flightlesstux/prompt-cachingupdated this year
Automatic prompt caching for Claude Code. Cuts token costs by up to 90%.
npm MIT -
Local observability for Codex sessions: TTFT, prompt-cache hit rate, token usage, per-step tool latency, and reasoning summaries (CLI + dashboard + MCP App panel).
MIT ★ 1 -
tokenwise-mcpupdated this year
Library + MCP server for reducing LLM token costs via compression, summarization, chunking, and Anthropic prompt-cache breakpoint planning
npm SEE LICENSE IN LICENSE -
Cross-session project memory for DeepSeek Harness - MCP server + archive daemon + panel. Local-first SQLite store with handoff packs tuned for prompt-cache prefixes.
MIT ★ 1 -
Sales assistant with Claude tool use, hybrid RAG, prompt caching and an MCP server, measured one change at a time with evals
-
@tokenzero/mcpactive
TokenZero MCP server (stdio): audit LLM token waste locally, price model alternatives, check workspace spend records and make prompts cache-friendly from Claude Code, Cursor, Windsurf or Cline.
npm Apache-2.0 -
Keeps LLM prompt-cache prefixes stable so provider prompt caching actually hits: local proxy (point your client's baseURL at localhost) + MCP server. Fully local, no telemetry.
npm Apache-2.0 -
Thinklysgithub.com/yotam4249/Thinklysupdated this year
RAG-based learning platform with a TypeScript agentic layer: custom MCP server exposes user-scoped tools over uploaded documents; a Claude tool-use agent drives them, with prompt caching, per-run traces, and a baseline-vs-agent evaluation harness.
-
mcp-osp-promptgithub.com/mcp-community/prompt-fetcherupdated this year
MCP server for fetching and caching prompt templates from GitHub, GitLab, or local directories
npm MIT -
FastAPI service that routes each LLM request to a provider by prompt complexity and logs tokens and cost to Supabase, with a semantic cache and an MCP server.
MIT -
atlas-agent-intelligence-layergithub.com/adrianofratelli-glitch/atlas-agent-intelligence-layeractive
AI agent intelligence layer in MongoDB Atlas: prompts, model config, semantic cache, guardrails and short/long-term memory stored as documents and changed live, plus an MCP tool-use loop.
MIT -
Auragithub.com/IzzoSol/Auraactive
Cut LLM token spend with zero dependencies. AURA answers recurring prompts free from cache, compresses conversation history, and distills bloated system prompts. CLI, MCP server (Claude, Cursor, Claude Code) and library.
npm MIT ★ 3 -
kakeibogithub.com/siash1/kakeiboupdated this year
A personal finance agent built from scratch on the raw Gemini API: hand-written tool-calling loop, double-entry PostgreSQL ledger, context caching, MCP server, 46-task eval harness, prompt-injection suite.
MIT -
llm-guard-gatewaygithub.com/ShreyAC962/llm-guard-gatewayupdated this year
Governance gateway for LLM and agent traffic: blocks prompt injection, redacts PII, enforces per-tenant token budgets, and serves repeat prompts from a semantic cache, cutting served p50 from 44.2 ms to 4.6 ms in the included benchmark. REST + MCP surfaces, offline-first local profile.
MIT -
LLM-nexusgithub.com/saanvijay/LLM-nexusupdated this year
A transparent HTTPS proxy between any AI client and upstream LLMs — redacts PII, compresses prompts, caches semantically similar responses, and surfaces everything in a real-time dashboard. Configurable via a single JSON file with upstream proxy chaining and MCP server integration.
MIT -
Deterministic context compressor for MCP agents. Slots in front of prompt caching, mem0, Letta, or Claude Code. Single Rust binary, zero ML at query time.
npm MIT ★ 9 -
⚡ Terminal HUD, hard budget breakers & prompt-cache leak alerts for AI coding agents (Claude Code, Cursor, Codex)
MIT ★ 32 -
Zero-dependency preflight CI linter, GitHub Action, SARIF analyzer & MCP server for CLAUDE.md, AGENTS.md, SKILL.md, .mcp.json & prompt caches
MIT -
Measure what context compression actually costs an LLM agent: where a session's tokens go, and whether a rewrite pays for the prompt cache it invalidates. Rust crate, MCP server and benchmark harness.
npm MIT ★ 1 -
korea-persona-interviewgithub.com/binaryloader/korea-persona-interviewupdated this year
Synthetic Korean persona interview tool over NVIDIA Nemotron-Personas-Korea (CC BY 4.0). CLI / MCP server / MCP orchestrator entry points, multi-provider LLM backend (OpenAI / Anthropic / OpenAI-compatible local), multi-turn interviews with auto follow-up, persona drift detection, and prompt caching.
MIT -
agentic-dev-assistantgithub.com/SaifulJnU/agentic-dev-assistantupdated this year
Multi-agent AI assistant that turns a GitHub issue into a draft PR with tests; planner/coder/reviewer loop on LangGraph, served via FastAPI + MCP. Swappable Claude / GPT / local Hugging Face models. Prompt caching & model tiering for ~50% lower cost. Dockerized, CI-tested.
-
Self-hosted LLM gateway that turns flat-rate AI subscriptions (Claude Code Max, ChatGPT Plus, Codex, Copilot, M365, Gemini) into one OpenAI/Anthropic-compatible API. Unified subscription wallet, OAuth passthrough on /v1/responses, prompt-injection + PII defense, MCP server, semantic cache, time-travel replay.
Apache-2.0 ★ 15 -
@yurdeth/image-gen-mcpactive
Multimodel image generation MCP server (OpenAI GPT Image + Google Gemini/Nano Banana). One tool: generate_image (text-to-image and editing via input_paths). Saves images to disk and returns the absolute path — never inline base64. Prompt-hash cache avoids
npm MIT -
Local cost analyzer for AI coding agents: see where tokens and AI Credits go in Claude Code, GitHub Copilot Chat, pi and mitmproxy sessions. Turn-by-turn cost, prompt caching explained, cost by repository and project, replay on another model, tool-call audit, MCP server with interactive views. Also a VS Code extension: AI Credits Analyzer.
MIT ★ 1 -
SEO Performanceno update data
Post-publish SEO performance MCP. Unifies **Google Search Console, Matomo, GA4, Microsoft Clarity, and AI-citation signals** per URL and emits a **verdict per post** (refresh / expand / merge / kill / double_down / hold) with deterministic reason codes and a 0-1 confidence score. Reporting only — never mutates posts. Works with any CMS via XML sitemap; optional Ghost Admin API integration for richer metadata. ## Tools - **posts.list** — discover posts via XML sitemap, JSON override, Ghost Admin API, or explicit URL list. - **posts.snapshot** — unified 30/60/90-day signal pull across every configured source for one URL. - **posts.decay_curve** — weekly GSC bucket + decay / plateau / growth classification. - **posts.verdict** — rule-based verdict engine: refresh, expand, merge, kill, double_down, hold. - **posts.refresh_brief** — markdown brief: verdict + reasons + numbers + suggested actions. - **posts.cite_loss** — list of LLMs that previously cited this URL but no longer do. - **cohort.report** — verdict engine across a cohort, ranked by verdict priority + confidence. - **gsc.quick_wins** — pages at positions 5-15 with low CTR, ranked by impressions. ## Prompts - **audit_cohort** — run cohort.report and generate refresh briefs for every actionable URL. - **find_quick_wins** — propose query-verbatim title and H1 rewrites for SERP underperformers. - **citation_loss_sweep** — find URLs that lost LLM citations and recover them. ## How it works Each source is best-effort: if its env vars are missing the field is omitted. Returns whatever is available. Optional DuckDB cache layer for historical comparisons. Platform-agnostic — works with WordPress, Hugo, Astro, Next.js, Webflow, Framer, Wix, Squarespace, Substack, Ghost, and any CMS that exposes a sitemap.
smithery -
cache-manager-mcpgithub.com/datawithdanny/cache-manager-mcpupdated this year
Save tokens across AI agent chats with prompt TTL tracking, handoff memories, and cost insights — over MCP.
MIT -
scorchmarkgithub.com/nas01010101/scorchmarkupdated this year
Find the cache tax draining your AI bill: a cross-provider cache-TTL-waste detector + model-swap savings simulator + pricing-drift + per-agent attribution. MCP server + CLI.
pypi MIT -
⚡ Claude Adaptive Token Optimizer — Zero-dependency, cache-preserving token optimization toolkit for Claude Code CLI and Desktop
MIT ★ 2 -
Aether Pulse — Open-Source AI Token Efficiency Leaderboard, Cache Doctor MCP Server & Architect Loadout Platform by Eloquent (@mjeanhudson-dev)
npm -
MCP server that lets Claude Code delegate bounded coding tasks to Kimi K2.7-code, optimized for Moonshot prefix caching
MIT -
@nomadop/session-watchergithub.com/nomadop/session-watcherupdated this year
Local Claude Code context-cost monitor, transcript replay, buckets, and handoff
npm MIT -
@sunit24blr/brain-mcpgithub.com/sunitlabsorg/interceptor-brainupdated this year
MCP server that acts as a project brain with handoffs, tasks, decisions, and phase rules.
npm MIT -
LLMxRay — Local LLM Observatory. Full observability interface for Ollama: chat, streaming, reasoning, RAG, introspection, metrics.
npm Apache-2.0 ★ 29 -
wakegithub.com/Fzhiyu1/wakeupdated this year
Substrate memory: full-KB injection into long-context LLM. Recall as shape activation, not retrieval. Three integration paths (lens plugin / MCP / hook).
NOASSERTION -
Measure what each successful task of your AI app costs — and prove when a change makes it cheaper without hurting quality. CLI, MCP server and Claude Code skill.
MIT -
A股多Agent量化交易决策框架 — LangGraph 15位AI分析师辩论式分析,反思闭环+Token优化,支持DeepSeek/Qwen/GLM,附零依赖MCP server | Multi-agent A-share trading framework with reflection loop & MCP server
MIT ★ 4 -
Request-path token proxy + analyzer for agent conversations. Live: cut 74% of input tokens on a real DeepSeek call (mcp-token-saver-proxy.fly.dev). MIT core, no-LLM, crypto-native Pro.
-
glyph-compressgithub.com/Neolambo/glyph-compressupdated this year
Cut the tokens your IDE-to-LLM coding session is billed for. Stops re-sending files the model already has, places provider cache breakpoints where they cover the request, and compresses what needs sending — measured against real tokenizers, not character counts. CLI, VS Code, MCP server, proxy.
npm AGPL-3.0 ★ 4