Claude Code is sharp, but it is blind. It cannot see your screen, control your browser, run a local model, transcribe audio, move your windows, or record a demo video. It writes great code and then asks you to do everything else manually.
MCP fixes that. Model Context Protocol is Anthropic's open standard for connecting AI assistants to external tools. And ToolPiper is a native macOS app that exposes over 300 of those tools from a single install.
This is the walkthrough. Install ToolPiper, connect it to Claude Code, and start using tools that actually touch your system.
What does Claude Code get from ToolPiper?
Over 300 tools organized into nine categories. Not just variations of the same thing. Nine distinct capability domains that cover most of what a developer does outside the editor.
| Category | Tools | What it does |
|---|---|---|
| Core AI | 11 | LLM chat, TTS, STT, embeddings, OCR, model lifecycle |
| Browser | 17 | CDP automation using accessibility tree selectors, not CSS |
| System | 29 | Window management, audio, display, Bluetooth, Dock, Spaces, Finder, and 22 more macOS domains |
| Testing | 6 | PiperTest: AX-native visual tests with self-healing and Playwright/Cypress export |
| Video | 16 | Screenplay, recording, rendering, narration, timeline editing |
| Outreach | 12 | GitHub, Hacker News, Reddit, X/Twitter, content queue |
| Analysis | 8 | Image/text analysis, RAG collections, web scraping, upscaling |
| Motion | 5 | Pose estimation and 60fps skeleton streaming via Apple Vision |
| OAuth | 4 | Connection management for external services |
Plus 5 MCP resources (server status, model catalog, backend health, saved tests, tool categories) and 5 workflow prompts for guided multi-step tasks.
How do you set it up?
Two steps. Total time: about 30 seconds if you already have Claude Code.
1. Install ToolPiper. Download from modelpiper.com or the Mac App Store. Open it once. A starter model downloads automatically. ToolPiper installs a symlink at ~/.toolpiper/mcp pointing to the native MCP binary bundled inside the app.
2. Connect to Claude Code.
claude mcp add toolpiper -- ~/.toolpiper/mcpRestart Claude Code. Done. You now have over 300 tools. No npm, no Docker, no Python, no API keys for local inference.
For Cursor, Windsurf, Gemini Code, or any MCP client that uses a JSON config file:
{
"mcpServers": {
"toolpiper": {
"command": "/Users/<you>/.toolpiper/mcp"
}
}
}ToolPiper also supports Streamable HTTP for web-based clients at http://localhost:9998/mcp.
What can you actually do with this?
Here are five real workflows, not hypotheticals. Each one runs entirely through Claude Code after connecting ToolPiper.
Run a local LLM without leaving the terminal
"Summarize this file using the local model." Claude Code calls chat, which runs the prompt through whatever model ToolPiper has loaded, llama.cpp on Metal GPU. No API key. No network request. The response comes back to Claude Code's context like any other tool result.
You can also call models to see what is downloaded, load_model to swap models, and transcribe to convert audio files to text. All on-device inference backed by Apple Silicon's Neural Engine and Metal GPU.
Automate Chrome without writing Playwright scripts
"Go to the staging site, click the login button, fill in the test credentials, and verify the dashboard loads." Claude Code calls browser_snapshot to read the page structure, then browser_action with accessibility-tree selectors like role:button:Sign In and label:Email. Each action returns an AX diff showing what changed on the page.
No CSS selectors to break. No test harness to configure. The tools auto-connect to Chrome on first call.
Record a demo video from natural language
"Create a 30-second demo video showing the signup flow, add narration explaining each step." Claude Code calls video_save to define the screenplay, video_record to capture the screen while executing browser actions, and video_narrate to add AI-generated voiceover. The result is an MP4 with narration, produced entirely from a text prompt.
Control your entire Mac
"Mute the speakers, set the display brightness to 50%, and move Slack to the left half of the screen." Claude Code calls action_audio, action_display, and action_window. 29 system tools cover windows, audio, display, Bluetooth, Dock, Spaces, Finder, Focus modes, notifications, calendar, contacts, reminders, power management, and more.
ToolPiper handles the execution. Every action runs through macOS APIs directly, not AppleScript.
Transcribe a meeting and query it
"Transcribe meeting.m4a, ingest it into the project knowledge base, then find every action item mentioned." Claude Code calls transcribe (on-device STT), rag_ingest to index the transcript, and rag_query to search it semantically. The entire pipeline runs locally. Your meeting audio never leaves your machine.
How is this different from stacking five MCP servers?
You could install Playwright MCP for browser automation, a filesystem server for file access, a separate STT service, and cobble together desktop control with shell scripts. People do this. It is painful for three reasons.
Context budget. Every MCP server dumps its tool definitions into the AI's context window. Five servers with 10 tools each means the AI is scanning 50 tool descriptions before deciding which one to call. ToolPiper's over 300 tools are organized with clear descriptions and tool annotations (read-only hints, destructive hints, idempotent hints) that help the AI choose correctly. Claude Code handles it well because Anthropic designed MCP with large registries in mind.
Composability. ToolPiper's tools share state. Transcribe audio, pass the text to the local LLM, query your RAG index for context, and automate a browser action based on the result, all through one server process. No inter-server coordination, no passing data between disconnected tools.
Zero maintenance. One app to update. ToolPiper auto-updates its MCP binary. No npm version conflicts, no Docker image pulls, no Python environment to manage.
What are the limitations?
macOS only, Apple Silicon required. ToolPiper's inference backends depend on Metal GPU, Neural Engine, and Core Audio. M1 or later. No Windows, no Linux.
Browser tools need Chrome. The CDP automation tier requires Chrome or Chrome Dev. Safari, Firefox, and Arc are not supported for browser automation.
Some features are Pro. Core tools are free: all inference (chat, transcribe, speak, embed, OCR), all browser automation, all read-only operations, full MCP access. ToolPiper Pro at $9.99/month adds test mutations, developer tokens, and advanced model management.
Over 300 tools is a lot of context. Claude Code handles it. Other clients may struggle with tool selection at this scale. ToolPiper supports category filtering (--category core for 11 tools, --category automation for 34) if you need to trim the set for a specific client.
Local models are not GPT-4. On-device inference with 3B-8B parameter models handles summarization, classification, code review, and most conversational tasks. Complex multi-step reasoning with large context windows still favors cloud models. ToolPiper also proxies to cloud APIs (OpenAI, Anthropic, Groq) when you want that.
What MCP clients work with ToolPiper?
Any MCP-compatible client. Tested with Claude Code, Cursor, Windsurf, and Gemini Code. The stdio transport works with CLI tools. The Streamable HTTP transport works with web-based clients. Tool definitions and behavior are identical across both.
For Claude Code, the experience is the best because Anthropic built MCP and their client handles large tool registries natively. Tool selection quality in other clients varies.