Toolpiper

io.github.ModelPiper由 GitHub 帳號 ModelPiper 發布v0.0.10-beta更新於 Oct 25, 2026

On-device MCP tools for macOS: local AI, speech, vision, browser automation, system control.

概覽

AI 產生的概覽

ToolPiper 是原生 macOS 應用程式,其本機 MCP 伺服器提供 450 多個裝置端工具,涵蓋本機 AI、語音、視覺、瀏覽器自動化與系統控制。

可以做什麼
  • 在 Mac 上執行本機大型語言模型對話、語音轉文字、文字轉語音與向量嵌入
  • 對圖片做 OCR、影像分析、文字分析,並檢索本機文件索引
  • 透過 CDP 驅動 Chrome:快照、操作、斷言、網路模擬、Web Vitals
  • 儲存、重播瀏覽器測試並匯出為 Playwright 或 Cypress
  • 控制系統:視窗、空間、剪貼簿、音訊、顯示器、Finder、行事曆、提醒事項、聯絡人、通知、程序、網路、藍牙與捷徑
  • 錄製與剪輯螢幕影片、比較 GitHub 儲存庫、管理內容佇列、讀取 Search Console,以及搜尋、下載和載入模型
適用情境
適合希望助理在 Mac 上直接操作本機並使用本機模型、不依賴雲端服務的情境,例如桌面自動化、瀏覽器測試、OCR 與語音、私有文件檢索。適用於執行 macOS 26 或更新版本的 Apple Silicon Mac。
執行需求
需要 macOS 26 或更新版本的 Apple Silicon Mac,並安裝 ToolPiper 應用程式且啟動一次。MCP 端點為回送位址 9998 埠上的 Streamable HTTP;回送連線無需驗證,區域網路與遠端連線使用 bearer 權杖。本機使用無需 API 金鑰或帳號,僅在下載模型或接入可選雲端服務時需要網路。
安裝前請注意
系統類工具可變更視窗、剪貼簿、檔案、行事曆、聯絡人與程序,瀏覽器工具會操作實際頁面,應確認允許助理執行哪些操作。此應用程式為閉源軟體,會為更新檢查、登入時的授權檢查與每日活動統計連線至其自有伺服器,並可選擇性向 PostHog 傳送使用計數(可關閉)。付費方案可解鎖更多工具。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 Toolpiper,將其新增到工作區。
  2. 為需要使用其工具的對話啟用該服務。

Desktop only,透過 STDIO。 STDIO 服務會啟動本機處理程序,因此需要 SourceWeft 桌面主機。

其他 MCP 客戶端

參照 儲存庫 中的啟動說明。

README

ToolPiper MCP Server for macOS

450+ Model Context Protocol tools, all running on-device. ToolPiper is a native macOS app that turns your Mac into a local AI platform: bundled inference (llama.cpp on Metal, Apple Intelligence, speech on the Neural Engine), an OpenAI-compatible API at localhost:9998, and a built-in MCP server that exposes the whole machine to any MCP client: Claude Code, Cursor, Codex, or your own scripts.

No cloud required for local models. No Docker. No Python. No API keys. Your prompts, files and outputs stay on your Mac, and you can check what ToolPiper connects to yourself: lsof -i -P | grep ToolPiper while it runs.

This repository is the official home for ToolPiper's MCP server manifest (server.json), the Claude Code plugin (plugins/toolpiper), and the setup guides. The ToolPiper app itself is not open source, but its privacy claims do not ask for trust: every byte of network traffic is observable on your own machine. Open-source siblings live in this org: PiperSR (ANE super-resolution model), FluidAudio, mlx-audio-swift.

Quick start

Install ToolPiper, launch it once, then connect any MCP client over Streamable HTTP. Loopback connections need no auth: no token, no OAuth step.

Download the disk image from modelpiper.com/download, or install with Homebrew:

bash
brew tap modelpiper/tapbrew trust modelpiper/tapbrew install --cask toolpiper

The trust line is required once: Homebrew 6 refuses casks from taps outside its own until you trust them.

Claude Code, as a plugin (one server entry, nothing to configure by hand):

bash
claude plugin marketplace add ModelPiper/toolpiperclaude plugin install toolpiper@modelpiper

Claude Code, as a manual server:

bash
claude mcp add -s user --transport http toolpiper http://127.0.0.1:9998/mcp

Codex, as a plugin:

bash
codex plugin marketplace add ModelPiper/toolpipercodex plugin add toolpiper@modelpiper

Or as a manual server: codex mcp add toolpiper --url http://127.0.0.1:9998/mcp. Codex signs in to ToolPiper once; approve ToolPiper's consent page when it appears.

Any other MCP client, in its server configuration:

json
{  "mcpServers": {    "toolpiper": { "type": "http", "url": "http://127.0.0.1:9998/mcp" }  }}

Claude Desktop, as a one-click extension: download toolpiper-1.0.1.mcpb from Releases and double-click it; Claude Desktop shows its install sheet. The JSON block above does not apply to Claude Desktop, whose config file runs local stdio servers only. Source, build script and privacy statement: desktop-extension/. It is a small Node.js forwarder that relays Claude Desktop's stdio to the same loopback server, and it serves the directory profile: the catalog minus every tool that uses an AI model to generate images, video or audio.

Installing with an AI agent? Point it at llms-install.md.

What's in the catalog

450+ tools, 5 resources, and 5 workflow prompts. Names follow <category>_<action> (vision_ocr, system_window_snap, browser_assert). The live catalog, with per-tool documentation, is at modelpiper.com/mcp-tools.

CategoryExamplesWhat it covers
Core AIchat, audio_transcribe, audio_speak, text_embedLocal LLM inference, speech-to-text and text-to-speech on the Neural Engine, embeddings
Apple frameworks + RAGvision_ocr, image_analyze, text_analyze, rag_queryApple Vision OCR, image analysis, natural-language analysis, hybrid HNSW + BM25 document search
Browser automationbrowser_snapshot, browser_action, browser_assert, browser_interceptChrome via CDP: AX-tree snapshots, self-healing selectors, network mocking, Web Vitals, coverage
Testingtest_run, test_save, test_exportSaved browser test sessions, fast replay, Playwright/Cypress export
System actionssystem_*Windows, Spaces, clipboard (with OCR'd history), audio, displays, Finder, Calendar, Reminders, Contacts, notifications, processes, network, Bluetooth, Shortcuts, and more
Video creatorvideo_record, video_render, video_edit_timelineScreen-recorded screenplays, non-destructive timeline editing, OTIO/FCP XML/EDL export
Outreachgithub_compare, queue_add, gsc_analyticsGitHub, the content queue, Google Search Console
Model managementmodel_search, model_download, model_loadSearch HuggingFace, download GGUF/CoreML, load/unload
Routing + integrationsendpoint_set, endpoint_recommend, claude_code_statusSwitch inference endpoints per session, see how Claude Code connects

Resources: toolpiper://status, toolpiper://models, toolpiper://backends, toolpiper://tests, toolpiper://categories Prompts: setup-model, audit-page, create-test, voice-over, research-topic

Category and profile filtering

Don't want the whole catalog in context? Pass category or profile as a query parameter on the /mcp URL. Both the tool list and the server instructions trim to match:

bash
# Just LLM inference (core category)claude mcp add -s user --transport http toolpiper "http://127.0.0.1:9998/mcp?category=core"
# Web dev workflow (automation profile: core + browser + testing)claude mcp add -s user --transport http toolpiper "http://127.0.0.1:9998/mcp?profile=automation"

Transport

TransportEndpointAuth
Streamable HTTPPOST http://127.0.0.1:9998/mcp (plus GET /mcp, the SSE channel for tools/list_changed)None for loopback peers, enforced at the kernel socket, so Host/Origin spoofing cannot widen it. LAN and remote peers present Authorization: Bearer.

Streamable HTTP is the only transport. ToolPiper has no stdio server, and older guides that described one are obsolete.

Privacy

Your prompts, files and outputs stay on your Mac when you use local models. ToolPiper contacts our servers for update checks, a license check when you're signed in and a daily activity count, and it can send optional usage counts. You can turn the last two off in Settings, and none of them include your content. There are no crash uploads, and local use needs no account. Everything else is traffic you ask for: model downloads from HuggingFace, a cloud provider you add, and tools that explicitly target the web. You can watch it yourself:

bash
lsof -i -P | grep ToolPiper

You will see update, license and activity requests to modelpiper.com hosts, optional usage counts to PostHog if you leave them on, plus any downloads or providers you chose, and none of them carry your content. Full policy: modelpiper.com/privacy.

Requirements and pricing

  • macOS 26 or later on Apple Silicon
  • Free: the full local model runner and every speech feature (transcription, text to speech, dictation). No caps, no account.
  • Paid plans add more of the catalog. The current ladder is at modelpiper.com/pricing.

FAQ

Is this the same as "ToolPipe"? No. ToolPiper (this page, modelpiper.com) is the macOS MCP server by ModelPiper. Unrelated projects with similar names exist in some directories.

What does the MIT license in this repo cover? The contents of this repository: the server manifest, the Claude Code plugin, and the setup guides. The ToolPiper app itself is proprietary.

Does it work without an internet connection? Yes. Inference, speech, vision, RAG, and system tools are all on-device. You only need the network to download models.

Can other machines connect? Loopback is zero-config. LAN and remote access go through bearer-token auth.

I already added ToolPiper with claude mcp add. Do I need the plugin? No. Claude Code keeps your manual toolpiper server and skips the plugin's copy while the manual one exists, so nothing doubles up. The plugin is for a fresh setup, or if you remove the manual entry.

來源:README.md,提交 5902f5b

工具

0
SourceWeft 尚未在沙箱中啟動過該軟體包,因此其工具未知。

版本歷史

2
  1. v0.0.10-beta最新Oct 16, 2026
  2. v0.0.9-betaSep 16, 2026

評分

已滿足適用檢查項的 50%

該評分衡量此收錄條目的完整程度,依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價,也不表示伺服器安全或目前可用:可連通性是帶日期的檢查結果,評分也不包含安裝量或使用者評價。

  • 近 30 天內可連通25 分

    請保持端點能夠回應 MCP 握手;SourceWeft 每 30 天探測一次。

  • 已宣告認證方式10 分

    請在註冊中心條目中宣告憑證請求標頭,或保持端點能夠回應探測,以便判定其認證方式。

  • 已列出工具15 分

    請讓伺服器在無憑證時回應 tools/list,或發布可在沙箱中無憑證啟動的軟體包。

  • 描述必需15 分
  • 圖示10 分
  • 文件15 分
  • 授權條款10 分