MacPaw AI
Contact

Mac-native intelligence

Explore MacPaw AI technologies
  • Secure
  • On-device
  • Mac-native

MacPaw AI is our applied-AI direction - initiatives that help Mac apps think, remember, and get things done right on the device. They bring inference, memory, and agents together in a native Swift runtime built with the aim of speed and privacy. We are implementing them in our products. Now, we're opening it to yours.

Built for Everyone interacting with MacPaw and its products - developers, product teams, and partners

Build with us

MacPaw AI Technologies

In development

Elix Mac-native agentic AI runtime

Learn about Elix

Build on-device AI for Mac in Swift. Describe what your app needs to do, and Elix handles the runtime - from execution and resource management to inference with Liquid Foundation Models.

  • Composable pipelines

    Write the intent in Swift. Elix works out what can run in parallel, so you can spend less time wiring AI infrastructure together. The pipelines layer is open source under Apache 2.0.

  • Shared service

    One service runs AI across apps and monitors resources. When the LLM isn't needed, Elix unloads it and frees the memory.

  • Inference engine

    Our custom MLX-based inference engine, tuned for Liquid Foundation Models and Apple silicon - running fully on the Mac.

In development

Mnemos On-device AI memory

Meet Mnemos

Give your AI a memory that stays on the Mac. It keeps context over time, knows where each fact came from, and is built with the aim of privacy by design.

  • More than a pile of memories

    People, projects, and facts stay connected instead of becoming disconnected chunks. When information conflicts, Mnemos can resolve it - without losing track of the source.

  • It knows how it knows

    Mnemos keeps apart what you said, what it found in its data, and what it observed. Each has its own trust level and traceable source.

  • Your memory goes with you

    Mnemos uses Portable Memory, an open, vendor-neutral format - export your memory, merge it, or move it between apps and devices.

Research preview

Computer Use AI that operates the Mac

Meet Computer Use

Give an agent the Mac itself. Computer Use is how AI sees the screen the way macOS does, acts in real apps, and gets measured on real Mac tasks - the research behind computer use in Eney.

  • Sees the screen like macOS does

    Screen2AX rebuilds a full accessibility tree from a screenshot, so agents and screen readers can use apps that never shipped one - with 2.2× better grounding than the native metadata.

  • Acts in real apps

    A native Swift agent clicks, types, and navigates Mac apps and the web one verified step at a time, and risky steps - paying, deleting, sending - wait for your approval.

  • Measured on real Macs

    MacArena scores agents on 421 human-verified tasks across 50 macOS apps, and GUIrilla-See, trained on tasks crawled from 1,108 real apps, posted the best macOS grounding score in the field at publication.

AI in our products

Eney is the first MacPaw product running on the stack. Here is what each technology does there, and where it goes next.

The local engine inside Eney - first of many

Eney is where the Elix stack comes together: reasoning runs as pipelines through one shared service, with Liquid Foundation Models inferred on the Mac through Elix's MLX engine - no round trip to the cloud for the work that can stay local.

Elix was designed as one service for many clients: a model loaded once and shared by every app on the Mac, unloaded when nobody needs it, with requests, caches, and resources managed together. Compact encoders catch sensitive data before it reaches a model; inference is tuned chip by chip, with the numbers published. Eney is the first client, not the last.

Learn about Elix
Visit Eney
  • Model 25.2Kdownloads

    macpaw-research/blip-icon-captioning

    BLIP-based captioner for macOS icons - drives accessibility labelling across our pipelines.

    View on Hugging Face
  • Model 16.5Kdownloads

    macpaw-research/yolov11l-ui-elements-detection

    YOLOv11-Large fine-tuned for macOS UI element detection - the backbone behind Screen2AX.

    View on Hugging Face
  • Model 15.1Kdownloads

    macpaw-research/yolov11l-ui-groups-detection

    Sibling YOLOv11-Large model - detects UI groups like toolbars and tab bars.

    View on Hugging Face
  • Dataset 5.3Kdownloads

    macpaw-research/UiPad

    UI Parsing and Accessibility Dataset - 352 macOS app screens with accessibility trees.

    View on Hugging Face
  • Dataset 4.3Kdownloads

    macpaw-research/mac-app-store-apps-metadata

    Structured metadata for thousands of Mac App Store apps.

    View on Hugging Face
  • Dataset 2.2Kdownloads

    macpaw-research/GUIrilla-Task

    Automated macOS UI exploration trajectories for GUIrilla-See.

    View on Hugging Face

Open source

Tools, ports, and engines we built for our own AI work and made available on GitHub.

GitHub
  • Swift 23stars

    MacPaw/ComposablePipelines

    SwiftUI-like DSL for composing AI pipelines in Swift.

    View on GitHub
  • Swift 2.9Kstars

    MacPaw/OpenAI

    Community-driven Swift package for the OpenAI API.

    View on GitHub
  • Swift 41stars

    MacPaw/Gliner2Swift

    Swift port of Gliner2 - entity recognition on the Mac.

    View on GitHub
  • Swift 1stars

    MacPaw/portable-memory-swift

    Swift implementation of the portable memory protocol.

    View on GitHub
  • Python 65stars

    MacPaw/macapptree

    Extract structured accessibility trees from any Mac app.

    View on GitHub
  • Python 10stars

    MacPaw/GUIrilla

    Automated agent for exploring and testing macOS GUI flows.

    View on GitHub
  • Python 12stars

    MacPaw/Fast-dLLM-mlx

    Fast-dLLM diffusion-LLM inference, ported to MLX.

    View on GitHub
  • Python

    MacPaw/portable-memory

    An open, vendor-neutral memory format and protocol for AI.

    View on GitHub

Research

Researching how AI can work better on the Mac - from portable memory and computer-use agents to AI pipelines and code evaluation.

Visit MacPaw Research

Publications

Browse all on MacPaw Research

Academic & research

We work with universities on AI research, internships, labs, and opportunities for students to learn by doing.

  • MIT

    Joint research and internships in on-device inference.

    Visit MIT

  • Kyiv-Mohyla Academy

    Joint applied-AI research and a steady ML talent pipeline.

    Visit KMA

  • Ukrainian Catholic University

    Joint research, internships, and a talent pipeline.

    Visit UCU

  • KPI Igor Sikorsky Polytechnic

    Home of our "Bilka Space" AI Lab and an ML scholarship.

    Visit KPI
The MacPaw AI team

Strategic partners

We work with specialists across foundation models, voice AI, and model training to bring the right technology into MacPaw AI initiatives.

  • Foundation models partner

    Liquid AI

    Liquid AI builds efficient Liquid Foundation Models. Elix is designed to run them even leaner on Apple silicon.

    Visit Liquid AI
  • Voice AI partner

    Respeecher

    Respeecher brings studio-grade voice cloning and real-time text-to-speech to Mac voice features through Elix.

    Visit Respeecher
  • Adapter training partner

    distil labs

    distil labs helps us train purpose-built small language models that can handle specific tasks without an expensive LLM call.

    Visit distil labs
Talk to the team

Build with MacPaw AI

Working on AI for the Mac? Bring us the product, technical problem, or idea you're exploring. We can talk about platforms, partnerships, model licensing, and roadmap.

Email the team [email protected]

Press about MacPaw AI

See what the press is saying about our AI products, platform, partnerships, and plans.