Skip to content
Free Direct Download

Everything Ollama does, free. Then it does the rest of your Mac.

Model downloads, the native llama.cpp engine, multi-model, a local OpenAI-compatible API. Free, no account, no caps, no terminal. Then over 450 MCP tools, browser automation, voice, and vision on top.

Powered by llama.cpp: ToolPiper embeds upstream llama-server directly, not a fork, currently build b12359, with the exact version shown in the About panel. Models are standard GGUF files. One install replaces Ollama + Open WebUI + Playwright MCP + a filesystem MCP + LangChain.

By , Founder & Lead EngineerUpdated
9 Backends

Local AI Inference

llama.cpp on Metal GPU

Run any GGUF model. Qwen 3.5, Llama 3.2, DeepSeek R1. 30+ tok/s on M2 Air. Flash attention, speculative decoding, and Jinja templates out of the box.

Apple Intelligence

Neural Engine inference. Summarization, rewriting, Smart Reply. Zero GPU impact, runs on dedicated ML hardware alongside your other models.

FluidAudio STT/TTS

Whisper-class transcription on your Mac. PocketTTS with zero GPU impact. Both run on the Neural Engine, leaving Metal GPU free for LLM inference.

MLX Audio TTS

Soprano, Orpheus, voice cloning via Qwen3 TTS on Metal GPU. High-quality voice synthesis with support for multiple speakers and styles.

Apple Vision OCR

On-device text extraction from images and PDFs. No model download needed, uses the Vision framework built into macOS.

Resource Intelligence

Real-time RAM monitoring via proc_pid_rusage. Automatic model eviction under memory pressure. Your Mac stays responsive without manual intervention.

Works with Claude Code, Cursor, Devin

Over 450 MCP Tools

Core AI (33 tools)

chat, audio_transcribe, audio_speak, audio_voice_clone, text_embed, vision_ocr, model_load/list/search/download, voice_chat memory, endpoint management. All inference on Neural Engine and Metal GPU.

System Control (164 tools)

26 native macOS domains, all in-process. Windows, displays, audio devices, network, Bluetooth, Dock, Spaces, Finder, calendar, contacts, focus, clipboard history, notifications, processes, reminders, timers.

Browser & Web (32 tools)

AX-native selectors, self-healing, 7 assertion types, visual recording, network interception, coverage, WebAuthn, autofill, plus web_scrape, http_request, youtube_transcript, and web_api_discover.

Video Creator (65 tools)

AI-driven screenplay-to-MP4 pipeline. screenplay, rehearse, record, render, narrate, plus full timeline edit (composition, narration, screenplay, timeline, export), import_media, and clip library.

Filesystem & Git (24 tools)

file_read, file_write, file_create, file_delete, file_list_directory, file_pick_directory, code_search, workspace_search, plus git_status, git_diff, git_commit, git_log, git_push, git_checkout.

Outreach & Social (11 tools)

github_repo_list, github_activity, github_compare, gsc_analytics, gsc_inspect, queue_publish, queue_add, queue_list. Build your distribution loop.

Analysis & RAG (48 tools)

rag_ingest, rag_query, rag_collection_list with local embeddings + HNSW vector index. image_analyze, text_analyze, image_upscale, video_upscale, image_transform, pdf_extract, qr_generate.

Capture & Motion (8 tools)

vision_screenshot, vision_color_pick, audio_record, plus 60fps pose streaming: pose_detect, pose_format_list, pose_stream_start, vision_stream_start/stop.

Testing & Integrations (14 tools)

PiperFlow (flow_save, flow_run, flow_list, flow_get, flow_export, flow_delete) with self-healing selectors. Plus OAuth (4), Sieve content filters (4) for end-to-end workflows.

Drop-in Replacement

OpenAI-Compatible API

/v1/chat/completions

Drop-in replacement for OpenAI SDK. Change base_url to localhost:9998, that's it. Streaming and non-streaming. Works with LangChain, LlamaIndex, Continue.dev, Aider, and anything that accepts a custom OpenAI base URL.

/v1/embeddings

Local vector embeddings for RAG pipelines. EmbeddingGemma on the Neural Engine (zero-setup, 768-dim) or llama.cpp embedding models. Content-addressed cache for repeat queries.

/v1/audio/speech

Text-to-speech synthesis. Three engines (FluidAudio, MLX Audio, PocketTTS), eight voices, all on-device. Same API format as OpenAI's TTS endpoint.

/v1/audio/transcriptions

Speech-to-text transcription on the Neural Engine. Parakeet V3 model. Same API format as OpenAI's Whisper endpoint.

Cloud API Proxy

Route cloud requests through ToolPiper with Keychain key injection. Your API keys never appear in code or .env files. One base URL handles both local and cloud models.

Developer Tokens

tp_<64hex> Bearer tokens for team sharing and CI pipelines. SHA-256 hashed, stored in macOS Keychain. Works as api_key in any OpenAI SDK.

AX-Native

Browser Automation & Testing

AX-Native Selectors

Query Chrome's real accessibility tree via CDP's Accessibility.queryAXTree. Not a DOM simulation, not injected JavaScript. The actual computed AX tree that screen readers consume.

Self-Healing

3 modes: passive, fuzzy AX match, AI-assisted. Broken selectors repair in 5-15ms with zero external calls. Free, not a paid add-on.

Visual Recording

Browse your app normally. Every interaction becomes an AX-enriched test step with element metadata, page context, and a mutation diff showing what changed.

Export to Playwright/Cypress

Deterministic, idiomatic code for CI. AX selectors map to each framework's native format: role:button:Sign In becomes page.getByRole('button', { name: 'Sign In' }).

On-Device

Voice AI & Media

Speech-to-Text

Parakeet V3 runs on the Neural Engine. Whisper-class accuracy with zero GPU impact. Transcribe meetings, lectures, and voice memos locally.

Text-to-Speech

Three engines, eight voices, all on-device. FluidAudio on Neural Engine for low-latency. MLX Audio on Metal GPU for studio-quality synthesis.

Voice Cloning

Clone any voice from a 10-second sample via Qwen3 TTS on Metal GPU. Create custom voices for narration, accessibility, or creative projects.

Video Upscale

PiperSR at 44.4 FPS on Apple Neural Engine. Double-buffered ANE+Metal pipeline. Real-time 2x upscale from 360p to 720p.

Push-to-Talk

Right Option = dictate anywhere (STT to paste). Right Command = voice command (STT to LLM with tool definitions to ActionRouter). System-wide hotkeys.

Web Scraping

CDP-based scraper using a real browser. 16-framework detection, readiness-aware extraction, 7 output formats: markdown, text, readability, AX tree, HTML, links, screenshot.

ToolPiper vs The Alternatives

ToolPiperOllamaLM StudioOpen WebUI
SetupSigned DMG, one drag to ApplicationsHomebrew + CLIDMG downloadDocker + Ollama required
Inference enginellama.cpp + 8 other backendsllama.cpp onlyllama.cpp + MLXNo engine (proxies to Ollama)
MCP toolsover 450 tools over Streamable HTTPNoneMCP client only (serves none)None
Browser automation14 CDP tools with AX selectorsNoneNoneNone
Voice AISTT + TTS + voice cloningNoneTranscription in the Bionic appNone
Resource monitoringReal-time RAM, automatic evictionNone (reads RAM once at startup)Memory fit shown in model searchNone (most-requested feature)
API compatibilityOpenAI-compatible (chat, embed, speech, transcribe)OpenAI-compatible (chat, embed)OpenAI- and Anthropic-compatible, native RESTWeb UI only
Cloud proxyKeychain key injection, one base URLNoneOwn hosted cloud (Bionic plans)None
TestingPiperFlow with self-healing + Playwright/Cypress exportNoneNoneNone
PriceFree; Pro $10/mo, Studio $29/mo, Max $49/moFreeFree locally; paid Bionic plans at $20 and $100 a monthFree

How It Works

1

Install

Download the signed DMG from modelpiper.com/download. Drag to Applications. Launch. Home recommends a model for your Mac's memory.

2

Connect

Open ModelPiper in your browser, or run claude mcp add -s user --transport http toolpiper http://127.0.0.1:9998/mcp for MCP.

3

Build

Chat, transcribe, automate browsers, run tests, build agents, all on localhost.

Your Content Stays Local

Local models run on your Mac's GPU and Neural Engine, and your prompts, files and outputs stay on your Mac. ToolPiper contacts our servers for update checks, a license check when you're signed in, a daily activity count and optional usage counts. You can turn the last two off in Settings, and none of them include your content.

Powered by llama.cpp

The inference engine is upstream llama-server, embedded directly, not a fork. Currently build b12359; the exact version ships in the About panel.

9 Backends, 1 App

llama.cpp, Apple Intelligence, FluidAudio, MLX Audio, Apple Vision OCR, CoreML upscale, coordinated by one process.

Open Standards

OpenAI-compatible API, Model Context Protocol, GGUF models, Playwright/Cypress export. No proprietary lock-in.

Simple Pricing

The engine, the MCP host and your voice are free for everyone: no account, no caps. Pro adds our native tools, Studio adds media production, and Max adds developer tools.

Free

Free

forever

  • Native llama.cpp engine, runs any GGUF model
  • Unlimited model downloads, multi-model switching
  • Local OpenAI-compatible API + embeddings
  • MCP host: install any MCP server and run its tools through ToolPiper
  • All speech: transcription, text-to-speech, voice cloning, dictation, voice chat
  • Talk to your Mac: spoken commands across 26 system domains
  • Local chat with tools, on any model you run
  • Apple Intelligence on the Neural Engine
  • Browser basics: connect to Chrome, snapshot a page, list tabs
  • Search Console reports: what your site already ranks for
  • Free companion apps (VisionPiper, AudioPiper, MediaPiper)
Try ModelPiper

Pro

$10/mo

  • Everything in Free
  • Browser automation: act, fill, extract, record, intercept, assert and run Autopilot loops in Chrome
  • Files, Git and archives: the agent reads, edits, searches and commits
  • Clipboard cache tools: the agent reads and writes your clipboard history
  • Mac automation: the agent drives apps, types and clicks, arranges windows and Spaces, launches and stops processes, writes defaults and undoes
  • Local RAG over your files
  • Web scraping and YouTube transcripts
  • Screen and image analysis: screenshot, OCR, color pick, webcam, live stream and capture region controls
  • Image composition: import, fill, outpaint, free transform, perspective, gradients and text
  • Agent speech and recording: record, transcribe, speak and clone a voice from a tool call
  • Cloud AI models in chat, voice and the API (bring your own keys)
  • Outreach toolkit: SEO tools (SERP capture with AI Overview detection, keyword discovery, rank check), GitHub activity, content queue
  • API discovery toolkit + request replay
  • Save captured video streams to disk
Get Pro

Studio

$29/mo

  • Everything in Pro
  • Image editing and generation: adjust, crop, rotate, flip, transform, cut out, inpaint, remove, segment, generate
  • Image upscaling (on-device)
  • Image workbenches: layered canvases, masks, adjustment layers
  • Storyboards and video: shots, narration, timeline, render
  • Video upscaling (ANE-native)
  • Image to video, Gaussian splats and 3D worlds
  • Screen recording, media extraction and hosting for the clips you make, plus live stream recording with rewind
  • Audio waveforms and silence detection
  • Pose detection (60fps streaming)
Get Studio

Max

$49/mo

  • Everything in Studio
  • Code (agentic AI code editor)
  • PiperFlow (self-healing browser flows)
  • Automations: run any tool, command or chat in a project automatically, once, on an interval or on a cron schedule
  • Priority support
Get Max

Frequently Asked Questions

Do I need Ollama?

No. ToolPiper embeds the same llama.cpp engine Ollama wraps, directly rather than as a fork, and the whole runner is free: unlimited GGUF downloads, multi-model switching, the local OpenAI-compatible API, no account. Models are stored as standard GGUF files you can use with any llama.cpp tool, not a proprietary blob format. If you keep Ollama installed, ModelPiper can still connect to it as a backend, but there's nothing it provides that ToolPiper doesn't run natively.

Will my Ollama tools keep working?

Yes - ToolPiper serves the Ollama API itself. The compatibility listener is on by default and claims localhost:11434 whenever that port is free (Settings → General), so anything that talks to localhost:11434 talks to ToolPiper: model list, streamed chat, embeddings, pulls, deletes, all on the native engine. It's served as the legacy dialect - every response carries a standards-based deprecation header pointing at the first-party /v1 API, which is where new integrations should land.

What models can I run?

Any GGUF model from HuggingFace (135,000+ available). ToolPiper also ships curated presets tested on Apple Silicon: Qwen 3.5, Llama 3.2, DeepSeek R1, Phi-4, Gemma 3, and more. Each preset shows exact RAM usage so you never download something your Mac can't run.

Does it work with Claude Code?

Yes. One command: claude mcp add -s user --transport http toolpiper http://127.0.0.1:9998/mcp. Restart Claude Code and all over 450 tools are available. Also works with Cursor, Devin, and any MCP-compatible client over Streamable HTTP at http://127.0.0.1:9998/mcp.

How much RAM do I need?

16GB minimum, the mainstream sweet spot (7B-8B models alongside normal app usage). 32GB opens up 14B models and multi-model workflows. ToolPiper's resource intelligence monitors RAM in real-time and automatically evicts models under memory pressure.

Is it really free?

Yes. The engine, the MCP host and your voice are free with no account and no caps: the native llama.cpp engine, unlimited model downloads, multi-model switching, the local OpenAI-compatible API, embeddings, developer tokens, all speech (transcription, text-to-speech, voice cloning, push-to-talk dictation, voice chat), Apple Intelligence, spoken Mac commands, and any MCP server you install. Our own native tools are the product: 207 of ToolPiper's 470 tools stay free. Pro ($10/mo) adds browser automation, files and Git, the clipboard cache, the agent driving your Mac (apps, typing and clicks, windows and Spaces, processes), local RAG over your files, web scraping with YouTube transcripts, screen and image analysis, the cloud API proxy for your own keys, the outreach toolkit, and API discovery and replay. Studio ($29/mo) is where the Mac makes things: image editing and generation, layered image workbenches, storyboards, video editing, image and video upscaling, and pose detection. Max ($49/mo) adds PiperFlow with self-healing, the agentic code editor, automation for every other feature, and priority support.

Does it work offline?

Yes, completely. All inference runs on your Mac's hardware. Once models are downloaded, everything works without an internet connection: chat, transcribe, speak, embed, OCR, browser automation, testing, and desktop control. The only features that require network access are cloud API proxy and the social/research tools (GitHub, Search Console, web search).

Ready to run AI locally?

One app. Nine backends. Over 450 tools. Zero configuration.