# ModelPiper

> ModelPiper is a suite of local-first AI apps for macOS. A browser-based web app connects to native companion apps that handle LLM inference, speech, vision, OCR, NLP, RAG, image/video upscaling, browser automation, and more, all running on-device with no cloud dependency.

## Products

- [ModelPiper Web App](https://modelpiper.com/chat): Browser-based local AI workspace. Chat with tools, image processing, and Studio, running against ToolPiper on your Mac or your own cloud providers. Chat is free to use; image editing and video upscaling require ToolPiper Studio ($29/mo) and E2E test authoring requires Max ($49/mo).
- [ToolPiper for macOS](https://modelpiper.com/download): Local AI gateway and MCP server. Bundles 9 inference backends (llama.cpp, Apple Intelligence, FluidAudio STT, FluidAudio TTS, MLX Audio TTS, Apple Vision OCR, CoreML Image Upscale, CoreML Pose, CoreML Embedding). 440 MCP tools across 14 categories: core inference, analysis, browser, filesystem, capture, media, testing, motion, outreach, search, sieve, system, video, and OAuth. Includes 26 system-action domains (audio, display, network, Bluetooth, calendar, contacts, Finder, notifications, power, process, reminders, system settings, window management) with 142 actions and push-to-talk voice commands (140ms STT on Apple Neural Engine). OpenAI-compatible API on localhost. Cloud API proxy with Keychain key injection. Free DMG download: the engine, the MCP host and your voice are free, our native tools are the product. Paid rungs: Pro ($10/mo) unlocks our tools: browser automation, files and Git, the clipboard cache, local RAG, web scraping with YouTube transcripts, screen and image analysis, the cloud API proxy, the outreach toolkit, and API discovery; Studio ($29/mo) adds image editing and generation, media projects, storyboards, video editing, image and video upscaling, and pose detection; Max ($49/mo) adds the agentic code editor and PiperTest. Team is $99/mo per deployment with unlimited named member tokens.
- [VisionPiper for macOS](https://modelpiper.com/visionpiper): Screen capture, MP4 recording, GIF conversion, and live WebSocket streaming. Region selection with pass-through interior. Free on the Mac App Store.
- [AudioPiper for macOS](https://modelpiper.com/audiopiper): Multi-source audio mixer and recorder. Captures microphone, system audio, and per-app audio via Core Audio Taps. Actor-based streaming WAV recording. Free on the Mac App Store.
- [MediaPiper Browser Extension](https://chromewebstore.google.com/detail/mediapiper/gknkckkachhflfmhdiaeaomacceechnb): Chrome, Firefox, and Safari extension for full-size image preview on hover. Declarative sieve engine resolves thumbnails to full-size images. Free to install; AI image upscaling runs through ToolPiper and requires Studio ($29/mo).

## Documentation

- [Homepage](https://modelpiper.com/): Product overview, features, pricing, and getting started guide
- [Documentation Hub](https://modelpiper.com/docs): All docs: 16 capabilities, 5 apps
- [ToolPiper API Reference](https://modelpiper.com/docs/toolpiper): REST API documentation, 90+ endpoints, dev tokens
- [MCP Server Docs](https://modelpiper.com/docs/mcp): 440 MCP tools, stdio + HTTP transports, setup for Claude Code / Cursor / Devin
- [MCP Landing Page](https://modelpiper.com/mcp): MCP tool categories, installation guide, per-feature filtering
- [MCP Tool Catalog](https://modelpiper.com/mcp/tools): Searchable catalog of all 440 built-in MCP tools: name, one-line description, category, and unlock tier for every tool
- [MCP Tools Marketplace](https://modelpiper.com/mcp-tools): Curated third-party MCP servers installable in ToolPiper with one click. 51 servers across developer tools, data, productivity, search, content, system, and AI tooling
- [GitHub MCP Server](https://modelpiper.com/mcp-tools/github): Create issues, manage pull requests, and search repositories from your AI assistant
- [Filesystem MCP Server](https://modelpiper.com/mcp-tools/filesystem): Read and write files in a directory you choose
- [Postgres MCP Server](https://modelpiper.com/mcp-tools/postgres): Run SQL queries against a Postgres database
- [Slack MCP Server](https://modelpiper.com/mcp-tools/slack): Read channels, post messages, and search Slack workspaces
- [Fetch MCP Server](https://modelpiper.com/mcp-tools/fetch): Fetch URLs and parse the content into markdown for the model
- [Git MCP Server](https://modelpiper.com/mcp-tools/git): Read, search, and manipulate Git repositories on disk
- [Memory MCP Server](https://modelpiper.com/mcp-tools/memory): Knowledge-graph backed persistent memory across chats
- [Sequential Thinking MCP Server](https://modelpiper.com/mcp-tools/sequential-thinking): Dynamic, reflective problem-solving through thought sequences
- [Time MCP Server](https://modelpiper.com/mcp-tools/time): Time and timezone conversion utilities
- [Everything MCP Server](https://modelpiper.com/mcp-tools/everything): Reference test server exercising every MCP feature
- [SQLite MCP Server](https://modelpiper.com/mcp-tools/sqlite): Query SQLite databases and run business-intelligence prompts
- [Redis MCP Server](https://modelpiper.com/mcp-tools/redis): Interact with Redis key-value stores from the assistant
- [Puppeteer MCP Server](https://modelpiper.com/mcp-tools/puppeteer): Headless browser automation and web scraping
- [GitLab MCP Server](https://modelpiper.com/mcp-tools/gitlab): Manage GitLab projects, issues, and merge requests
- [Google Drive MCP Server](https://modelpiper.com/mcp-tools/google-drive): Search and read files from Google Drive
- [Google Maps MCP Server](https://modelpiper.com/mcp-tools/google-maps): Geocoding, directions, and place lookup via Google Maps
- [Brave Search MCP Server](https://modelpiper.com/mcp-tools/brave-search): Web and local search via Brave's privacy-focused search API
- [Sentry MCP Server](https://modelpiper.com/mcp-tools/sentry): Retrieve and analyze issues from Sentry.io
- [AWS Knowledge Base MCP Server](https://modelpiper.com/mcp-tools/aws-kb-retrieval): Retrieve answers from an AWS Bedrock Knowledge Base
- [EverArt MCP Server](https://modelpiper.com/mcp-tools/everart): AI image generation using multiple model backends
- [ModelPiper Web App Docs](https://modelpiper.com/docs/modelpiper): Chat, providers, Studio, ToolPiper connection
- [VisionPiper Docs](https://modelpiper.com/docs/visionpiper): Screen capture, recording, streaming
- [AudioPiper Docs](https://modelpiper.com/docs/audiopiper): Audio mixer, per-app capture, recording
- [MediaPiper Docs](https://modelpiper.com/docs/mediapiper): Browser extension, sieve engine, AI upscale
- [Blog](https://modelpiper.com/blog): 152 technical articles and papers
- [Blog RSS Feed](https://modelpiper.com/blog/feed.xml): RSS 2.0 feed of all blog articles and papers
- [Workflows](https://modelpiper.com/workflow): 9 definitive guides to local AI workflows on Mac
- [Help](https://modelpiper.com/help): Support answers covering install, permissions, model downloads, offline behaviour and licensing. The same curated articles ToolPiper's assistant answers from, so the page and the app agree
- [Changelog](https://modelpiper.com/changelog): Versioned release notes, also available as Markdown at /changelog.md
- [OpenAPI Specification](https://modelpiper.com/api/openapi.json): Machine-readable OpenAPI 3.1 spec for the ToolPiper REST API (200+ operations)
- [API Catalog](https://modelpiper.com/.well-known/api-catalog): RFC 9727 linkset pointing at the spec and human-facing docs

## Model Fit and Mac Hardware

Which local AI models a given Mac can actually run, from measured memory bandwidth and unified memory ceilings rather than vendor claims.

- [Model Fit Checker](https://modelpiper.com/fit): Pick a Mac, see which of 3,641 open models fit, with quantization recommendations, memory requirements, and estimated tokens/sec.
- [Benchmark Leaderboard](https://modelpiper.com/fit/benchmarks): How prompt-processing and token-generation rates are measured by chip, quantization, and backend. Community results are not published yet; the per-model tok/s figures on /fit pages are formula estimates and are labelled as such.

### Per-model pages

The catalog carries **3,641 models, one page each**, at `/fit/<model-slug>`. Every
page answers the same question for one model: which Macs run it, at which
quantization, with what memory headroom and what tokens/sec.

Slugs are derived from the HuggingFace repo id: lowercased, with `/` and `.`
both replaced by `-`. So `Qwen/Qwen2.5-7B-Instruct` becomes
`/fit/qwen-qwen2-5-7b-instruct`, and any model id in the catalog can be turned
into its URL by that rule without looking it up.

Each page carries: parameter count, architecture, context length, license,
per-quantization memory requirements (Q4 through F16), a fit verdict for every
Apple Silicon chip, and estimated generation speed.

A representative head, chosen for coverage of size and task rather than raw
download rank (which is dominated by sentence-embedding models):

- [Qwen3 0.6B](https://modelpiper.com/fit/qwen-qwen3-0-6b): 0.75B general, runs on every Apple Silicon Mac including 8 GB
- [Qwen3 4B](https://modelpiper.com/fit/qwen-qwen3-4b): 4B general
- [Mistral 7B Instruct v0.3](https://modelpiper.com/fit/mistralai-mistral-7b-instruct-v0-3): 7.25B chat
- [Qwen2.5 7B Instruct](https://modelpiper.com/fit/qwen-qwen2-5-7b-instruct): 7.62B chat
- [Llama 3.1 8B Instruct](https://modelpiper.com/fit/meta-llama-llama-3-1-8b-instruct): 8.03B chat
- [Qwen3 8B](https://modelpiper.com/fit/qwen-qwen3-8b): 8.19B general
- [Qwen2.5 VL 7B Instruct](https://modelpiper.com/fit/qwen-qwen2-5-vl-7b-instruct): 8.29B vision-language
- [Qwen3.6 27B](https://modelpiper.com/fit/qwen-qwen3-6-27b): 27.78B multimodal
- [Qwen3 30B A3B](https://modelpiper.com/fit/qwen-qwen3-30b-a3b): 30.53B mixture-of-experts
- [Gemma 4 31B IT](https://modelpiper.com/fit/google-gemma-4-31b-it): 31B multimodal
- [DeepSeek R1 Distill Qwen 32B](https://modelpiper.com/fit/deepseek-ai-deepseek-r1-distill-qwen-32b): 32.76B reasoning
- [Qwen2.5 Coder 32B Instruct](https://modelpiper.com/fit/qwen-qwen2-5-coder-32b-instruct): 32.51B coding
- [Qwen3 32B](https://modelpiper.com/fit/qwen-qwen3-32b): 32.76B general
- [Llama 3.3 70B Instruct](https://modelpiper.com/fit/meta-llama-llama-3-3-70b-instruct): 70.55B chat, needs 64 GB unified memory at Q4
- [GPT-OSS 120B](https://modelpiper.com/fit/openai-gpt-oss-120b): 120.41B general, Mac Studio territory
- [MacBook Air](https://modelpiper.com/macbook-air): Which LLMs run on a fanless MacBook Air, M1 through M5.
- [MacBook Pro 14"](https://modelpiper.com/macbook-pro-14): Base, Pro, and Max chips compared by memory bandwidth and unified memory.
- [MacBook Pro 16"](https://modelpiper.com/macbook-pro-16): Pro and Max chips, M1 through M5, ranked for local inference.
- [Mac mini](https://modelpiper.com/mac-mini): The cheapest way into local AI on Apple Silicon.
- [Mac Studio](https://modelpiper.com/mac-studio): Max and Ultra chips, the highest memory bandwidth and unified memory Apple ships, up to 512 GB.
- [iMac](https://modelpiper.com/imac): Base M chips, where unified memory is the limit that matters.

### Per-chip pages

Each of the six products above also has one page per chip it ships, at
`/<product>/<chip>`, **42 pages** in total. The chip slug lowercases the
Apple chip name and replaces spaces with `-`, so `M3 Ultra` under Mac Studio
is `/mac-studio/m3-ultra`. Worked examples:

- [Mac Studio M3 Ultra](https://modelpiper.com/mac-studio/m3-ultra): highest memory bandwidth and unified memory Apple ships
- [MacBook Pro 14" M4 Max](https://modelpiper.com/macbook-pro-14/m4-max): the portable ceiling for local inference
- [Mac mini M4 Pro](https://modelpiper.com/mac-mini/m4-pro): cheapest route to 64 GB unified memory

Each chip page carries memory bandwidth, unified memory options, CPU/GPU core
counts, Neural Engine TOPS, the largest dense model that chip holds at Q4, and
a per-model fit table.

## Key Facts

- ToolPiper is an MCP server with 440 tools across 14 categories and 5 resources, stdio and HTTP transports
- 9 bundled inference backends: llama.cpp, Apple Intelligence, FluidAudio STT, FluidAudio TTS, MLX Audio TTS, Apple Vision OCR, CoreML Image Upscale, CoreML Pose, CoreML Embedding
- PiperSR video upscale: 44.4 FPS on M4 Max (1.5x realtime) using double-buffered ANE+Metal pipeline
- Custom CDP browser automation engine with AX-native selectors, self-healing (~5-15ms), and mutation diffing
- PiperTest (Max, $49/mo): visual AX-native test format with Playwright/Cypress export
- 16-framework detection in web scraper (React, Vue, Angular, Svelte, Next.js, Nuxt, etc.)
- HNSW + BM25 hybrid RAG with semantic chunking, EmbeddingGemma embeddings on the Neural Engine (zero-setup)
- Voice cloning via Qwen3 TTS with reference audio
- Push-to-talk dictation: 140ms on Apple Neural Engine, 25 languages
- ToolPiper system actions: 26 action domains, 142 actions, 181 MCP tools for desktop automation (built in to ToolPiper)
- All AI processing runs locally on-device, no cloud required unless user opts in
- ToolPiper requires macOS 26 or later, optimized for Apple Silicon
- Privacy-focused: prompts, responses, and workflows never leave the device
- All companion apps (VisionPiper, AudioPiper, MediaPiper) are free to install; MediaPiper's AI upscaling requires ToolPiper Studio ($29/mo)

## MCP Server Setup

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

Or in .mcp.json: `{ "mcpServers": { "toolpiper": { "type": "stdio", "command": "~/.toolpiper/mcp" } } }`

HTTP transport: `{ "mcpServers": { "toolpiper": { "type": "http", "url": "http://localhost:9998/mcp" } } }`

## Optional: llms-full.txt

For a detailed version with all articles and documentation pointers, see [llms-full.txt](https://modelpiper.com/llms-full.txt).

## Community

- [GitHub Discussions](https://github.com/ModelPiper/modelpiper-community/discussions): Community discussion board
- [Hugging Face](https://huggingface.co/ModelPiper): Published model weights, including PiperSR-2x
- [X / Twitter](https://x.com/modelpiper)
- [LinkedIn](https://www.linkedin.com/company/piperkit/)
- [Reddit](https://www.reddit.com/r/ModelPiper/)

## Legal

- [Privacy Policy](https://modelpiper.com/privacy)
- [Terms of Service](https://modelpiper.com/terms)