Local MCP Server on Mac: 93 AI Tools in One Install

Claude Code, Cursor, Windsurf, and other AI coding tools are only as powerful as the tools they can reach. MCP (Model Context Protocol) lets you extend them with custom tools. But most MCP servers do one thing. A file reader. A database connector. A web searcher. What if one server gave your AI assistant access to local inference, browser automation, voice, vision, web scraping, desktop control, and testing, all running on your Mac?

That is what ToolPiper does. It is a native macOS app that exposes 104 MCP tools across nine capability tiers. This article explains what MCP is, why local matters, and what those 104 tools actually do.

What is MCP and why should you care?

MCP stands for Model Context Protocol. It is an open protocol from Anthropic for connecting AI assistants to external tools. The architecture is simple: a server exposes tools (functions the AI can call), a client (Claude Code, Cursor, Windsurf) discovers those tools and invokes them, and the protocol handles JSON-RPC communication between the two.

Think of it as a USB port for AI. Before MCP, every AI assistant had its own plugin system, its own API format, its own integration story. MCP standardizes all of that. Write a tool once, and any MCP-compatible client can use it.

The ecosystem is growing fast. There are MCP servers for Playwright browser automation, filesystem access, database queries, GitHub, Slack, and dozens more. But most of them share a common limitation.

Why does it matter that an MCP server runs locally?

Most MCP servers are cloud wrappers. They proxy your requests to external APIs, which means your data still leaves your machine. The MCP server for OpenAI sends your prompts to OpenAI. The MCP server for a web search engine sends your queries to that search engine. The protocol is local, but the execution is not.

A local MCP server is different. The tools run on your hardware. Your data stays on your disk. When you ask your AI assistant to transcribe audio, summarize a document, or chat with a local LLM, the computation happens on your Mac's Neural Engine and Metal GPU. Nothing crosses a network boundary.

This matters for three reasons. Privacy: your code, your documents, your voice recordings never leave your machine. Speed: no round trip to a data center, no cold starts on someone else's infrastructure. Availability: it works on a plane, on a train, and when your ISP is down.

What is ToolPiper?

ToolPiper is a native macOS application that bundles multiple inference backends (llama.cpp, Apple Intelligence, FluidAudio, MLX Audio, Apple Vision OCR, CoreML) behind a single HTTP gateway on localhost. It is the local engine that powers ModelPiper, the visual AI pipeline builder.

It is also a full MCP server. One install gives your AI assistant access to 104 tools spanning local inference, browser automation, desktop control, testing, web scraping, video creation, and more. Setup takes 30 seconds.

How do you connect ToolPiper to Claude Code?

Install ToolPiper from modelpiper.com. Then run one command:

claude mcp add toolpiper -- ~/.toolpiper/mcp

Restart Claude Code. That is it. Your AI assistant now has 104 tools available. No npm install, no Docker, no Python environment. ToolPiper installs a symlink at ~/.toolpiper/mcp pointing to the native binary bundled inside the app. It updates automatically when you update the app.

ToolPiper supports two MCP transports. stdio is the universal one, used by Claude Code, Cursor, and most CLI tools. A lightweight Swift executable reads JSON-RPC from stdin and bridges to ToolPiper's local API. Streamable HTTP is the newer transport for web-based clients, served directly from ToolPiper's built-in HTTP server at POST localhost:9998/mcp. Both transports share identical tool definitions and handler logic.

What are the 104 tools?

The tools are organized into tiers by capability. Here is the full breakdown.

Tier 1: Core AI (8 tools). These are the foundation. chat runs prompts through a local LLM. transcribe converts audio to text using on-device speech recognition. speak synthesizes speech from text. ocr extracts text from images and PDFs using Apple Vision. embed generates vector embeddings for RAG. analyze_image and analyze_text run multimodal and text analysis. load_model loads a specific model into memory. All inference runs on your Mac's Neural Engine and Metal GPU.

Tier 2: Advanced AI (5 tools). models lists available models with download status. status reports server health and loaded backends. rag_collections and rag_query provide local retrieval-augmented generation over your documents. scrape extracts content from web pages in seven formats (markdown, text, readability, AX tree, HTML, links, screenshot) with framework-aware readiness detection for 16 JavaScript frameworks.

Tier 3: Browser Automation (14 tools). Full CDP-based browser automation using the accessibility tree instead of fragile CSS selectors. browser_snapshot captures the current page state. browser_action performs clicks, fills, and interactions with AX diff output showing what changed. browser_assert provides seven assertion types with polling. browser_console reads console output. browser_record captures interaction sequences. browser_manage handles connection lifecycle. browser_network monitors network traffic. browser_storage manages cookies and local/session storage. browser_performance measures Web Vitals. browser_coverage tracks JS/CSS code coverage. browser_eval executes arbitrary JavaScript. browser_intercept mocks network requests. browser_webauthn simulates passkey authentication. browser_autofill fills credit card and address forms.

Tier 4: PiperTest (6 tools). A visual, AX-native test format with self-healing selectors. test_list, test_get, test_save, test_delete manage test sessions. test_run executes tests and returns pass/fail results. test_export renders tests as Playwright or Cypress code.

Tier 5: Pose Detection and Streaming (5 tools). Real-time skeleton tracking using Apple Vision, with four output formats (compact binary, compact JSON, verbose JSON, rendered PNG). Pose detection and streaming tools for motion capture and animation workflows.

Tier 6: Scrape and Detect (2 tools). scrape extracts structured content from web pages using a real browser (not HTTP fetch). browser_detect identifies JavaScript frameworks and readiness state on the current page.

Tier 8: ActionPiper Desktop Control (29 tools). Full macOS system control. action_window manages windows (list, snap, move, resize, close, minimize, fullscreen, layouts). action_input simulates keyboard and mouse. action_audio controls volume and audio devices. action_display manages brightness and display settings. action_network handles Wi-Fi, DNS, and connectivity. action_bluetooth, action_dock, action_desktop, action_spaces, action_focus, action_media control their respective system domains. action_power, action_process, action_finder manage system resources. action_app_list, action_app_action, action_app_snapshot, action_app_assert interact with running applications. action_accessibility and action_appearance manage system preferences. action_notification, action_calendar, action_contacts, action_reminders access macOS services. action_system, action_storage, action_defaults, action_location, action_shortcut round out the system control suite.

Tier 9: Video Creator (12 tools). AI-driven video production. video_list, video_get, video_save, video_delete manage video screenplays. video_rehearse dry-runs a screenplay. video_record captures the screen during execution. video_render produces the final video. video_narrate adds AI-generated narration. video_preview shows a preview. video_edit_screenplay, video_edit_composition, video_edit_narration modify individual aspects of a video project.

Social and Research tools. github_activity and github_repos browse GitHub. hn_trending and hn_search pull from Hacker News. reddit_search and reddit_post access Reddit. x_browse and x_compose interact with X/Twitter. youtube_transcript fetches video transcripts.

How does this compare to other MCP servers?

Most MCP servers are single-purpose. Here is how ToolPiper compares to the alternatives you would need to cobble together for equivalent coverage.

The key difference is not just tool count. It is that ToolPiper's tools can compose. A single workflow can transcribe audio, pass the text to a local LLM, query your RAG index for context, and automate a browser action based on the result. All through one server process that shares model state and authentication.

What are the honest limitations?

ToolPiper is not for everyone. Here is what you should know before installing.

macOS only. ToolPiper is a native macOS application that requires Apple Silicon (M1 or later). It will not run on Windows or Linux. The inference backends depend on Metal GPU, Neural Engine, and Core Audio frameworks that are only available on macOS.

Some tools require ToolPiper Pro. The core tools are free. Test mutations (save, delete), developer tokens, and some advanced features require a ToolPiper Pro subscription at $9.99/month. All read-only tools and all inference tools are free tier.

Browser tools require Chrome. The browser automation tier uses Chrome DevTools Protocol (CDP). You need Chrome or Chrome Dev installed. Safari, Firefox, and Arc are not supported for browser automation.

104 tools is a lot. Tool selection quality depends on the AI client. Claude Code handles the large tool set well because Anthropic designed MCP with large tool registries in mind. Other clients may struggle with tool selection when presented with 93 options. ToolPiper's tool descriptions are written to help AI models choose correctly, but your mileage may vary across different clients.

Does this work with Cursor and Windsurf?

Yes. Any MCP-compatible client can connect to ToolPiper. For stdio-based clients like Claude Code and Cursor, use the symlink at ~/.toolpiper/mcp. For HTTP-based clients, point them to http://localhost:9998/mcp. The tool definitions and behavior are identical across both transports.

The configuration syntax varies by client. For Claude Code:

claude mcp add toolpiper -- ~/.toolpiper/mcp

For Cursor and other clients that use a JSON config, add ToolPiper to your MCP server configuration pointing to the same binary.

Do all 104 tools run locally?

All inference tools run locally. Chat, transcribe, speak, embed, OCR, image analysis, pose detection, and video upscale all execute on your Mac's hardware. Your prompts, audio, and documents never leave your machine.

The social and research tools (GitHub, Hacker News, Reddit, X, YouTube) make network requests to fetch public data. Browser automation tools interact with whatever page is loaded in Chrome, which may involve network traffic depending on the page. Desktop control tools operate on your local macOS system.

Is this free?

ToolPiper's free tier includes all inference tools (chat, transcribe, speak, embed, OCR), all browser automation tools, all read-only operations, and MCP access through both transports. You can use ToolPiper as a full MCP server without paying anything.

ToolPiper Pro ($9.99/month) adds test mutations, developer tokens, advanced model management, and priority features. The MCP server itself is not gated behind Pro.

How do you see what tools are available?

From Claude Code, you can ask your AI assistant to call the status tool, which returns server health, loaded models, and available capabilities. You can also call models to see which AI models are downloaded and ready to use.

ToolPiper also serves a full OpenAPI spec at http://localhost:9998/v1/openapi.json documenting all REST endpoints, and the MCP tool definitions include detailed descriptions and JSON Schema for every parameter.

For a deeper look at how the MCP server was built, including the two-transport architecture and the shared-definition pattern, see Building 41 MCP Tools in Swift (the tool count has since grown to 93, but the architecture remains the same).

This article is part of the local-first AI on macOS series. For the visual testing tools, see Visual Testing, No Code.