Labsco
MCP SERVER

ToolPiper

by ModelPiper

Turn a Mac into a local tool platform — inference, vision, browser automation and around 140 system actions, served over MCP with nothing leaving the machine.

Desktop & OS Automation
Summary
The catalog is the problem and the fix is in the same page.

A 300-tool server does not fit comfortably in most agents' context, which is why the category filter matters more than any single tool: connect with `--category core` for 11 tools, or `--profile automation` for 34, and both the tool list and the server instructions trim to match. Everything runs on the machine, and that claim is checkable rather than promised — `lsof -i -P | grep ToolPiper` shows what it is talking to while it runs.

What it is

A native macOS app that runs a built-in MCP server alongside on-device inference. It exposes the machine itself to any MCP client: local LLM chat, speech and vision on the Neural Engine, Chrome automation over CDP, screen recording, and system control across 26 macOS domains.

What you get
  • Local inference and speech — `chat`, `audio_transcribe`, `audio_speak` and `text_embed`, all on-device
  • Apple frameworks and retrieval — `vision_ocr`, `image_analyze`, `text_analyze` and `rag_query` over a hybrid HNSW and BM25 index
  • Browser automation shaped for agents — `browser_snapshot` returns an accessibility-tree snapshot, `browser_action` drives the page, `browser_assert` gives a verdict, `browser_intercept` mocks network responses
  • Saved browser tests you can replay and export — `test_run`, `test_save`, `test_export`
  • System control across windows, Spaces, clipboard, audio, displays, Finder, Calendar, Reminders, Contacts, notifications, processes, network and Bluetooth, under the `system_*` prefix
  • Video work — `video_record`, `video_render`, `video_edit_timeline`, with OTIO, FCP XML and EDL export
  • Model management — `model_search`, `model_download` and `model_load` against HuggingFace
  • Five resources including `toolpiper://status` and `toolpiper://models`, and five workflow prompts such as `audit-page` and `create-test`
Requirements

macOS on Apple Silicon. Install the app, launch it once, then connect over Streamable HTTP at `http://127.0.0.1:9998/mcp` — loopback needs no auth, and that is enforced at the kernel socket rather than by checking headers, so LAN peers still need a bearer token. A stdio path exists too through the `~/.toolpiper/mcp` symlink created at launch. The local model runner is free with no account; the Pro tier at $10/mo adds all nine inference backends and the full voice suite. This repository holds the server manifest and setup guides; the app itself is not open source.

Setup effort

One command — claude mcp add --transport http toolpiper http://127.0.0.1:9998/mcp