Silver desktop computer and laptops connected by a green cable. Hardware illustration.

Open source. Open possibilities.

MachBoost.

Local AI, on your terms.

Chat with your models. Code in your workspace.
Put your team's Macs to work, together.

Download for Mac Start in your terminal

Apple Silicon · macOS 14+ · Free community preview

Your workflow.
Already invited.

Claude DesktopClaude CodeCodexOpenAI APIOllama API

01 / A home for your models

From a quick question
to your next commit.

A native Mac app for the way you actually work. Download a compatible model, open a conversation, and go.

MachBoost 0.16.17 with a Muse Glimmer model, reasoning and repository tool timeline, reviewed code changes, and the composer effort selector. Sample data from a UI test. — Native macOS app 0.16.17 · sample workspace

A conversation that stays with you.

Streaming Markdown, reasoning controls, saved chats, and files that stay attached across follow-ups.

Give your model a workspace.

Read code, review tool calls, and inspect changes. Dev mode puts repository access and edit permissions within reach.

Bring a few more skills.

Connect MCP tools and custom instructions. Use text or vision models with capabilities checked before download.

Explore the desktop guide

Local chat historyBundled Python + MLXCompatible vision modelsNo telemetry uploads

02 / Better, together

One powerful Mac.
A whole team of possibilities.

Your teammates don't all need a Mac Studio. Connect to a shared host from MachBoost or point a compatible coding tool at its API.

Design MacBookMachBoost app

Engineering MacBookClaude Code

Team workstationOpenAI-compatible client

Authenticated private network

Primary hostStudioResident MLX models

Shared inference

Another available Mac

Optional compatible host or local fallback

Keep models ready.

Resident models, bounded queues, and employee limits keep a shared server manageable.

Choose where work runs.

The desktop app can prefer a device or choose a compatible host using residency and queue pressure.

Keep access deliberate.

Scoped keys for teammates. Repository tools stay on the client. Paid API fallback is an explicit choice.

Routing happens before generation; an active stream stays on its host. Generic API clients use the endpoint they target. Memory isn't pooled across devices.

Set up your team

03 / Less repeated work

Let the model think.
Skip what it already knows.

Exact reusable context can avoid expensive prompt processing. Here's where that made a measured difference.

Repository-prefix reuse

2.89×

median paired speedup

10 / 10 token-identical pairs

Same model. Same weights. Same prompts.

Lower time is better

Native MLX · cache off3.501 s

MachBoost · prefix reuse1.186 s

Qwen2.5 3B · 4-bit · Apple M5 Pro · 48 GB memory
10 repository questions · greedy generation · up to 48 output tokens
Recorded August 15, 2026 on MachBoost 0.11.0. Bars show medians; 2.89× is the median of paired ratios.

This measures repeated repository context, not faster token decoding. A first question, a new image, or entirely new context may get little or no speedup. These are historical results, not a performance guarantee for every model or this release.

Read the recorded results Run the benchmark All experiments, including limitations

04 / Bring your own workflow

Familiar commands.
Your own compute.

Start in the terminal, connect Claude Desktop, or build on the same API that powers the app.

OpenAI Chat Completions & Responses
Anthropic Messages
Ollama-compatible endpoints

Browse runnable examples

# Install the native text backend
python3 -m pip install "machboost[mlx] @ git+https://github.com/Machboost/Machboost.git@v0.16.17"

# Download a model, then start chatting
machboost pull qwen2.5:3b
machboost run qwen2.5:3b

# Run the API server
machboost serve

The Mac app bundles its own runtime. The CLI example installs into your Python environment.

import os
from openai import OpenAI

client = OpenAI(
    base_url="http://127.0.0.1:11435/v1",
    api_key=os.environ["MACHBOOST_API_KEY"],
)

for chunk in client.chat.completions.create(
    model="qwen2.5:3b",
    messages=[{"role": "user", "content": "Hello!"}],
    stream=True,
):
    if chunk.choices:
        print(chunk.choices[0].delta.content or "", end="", flush=True)

Install the OpenAI SDK, load the model, and set your host's API key. For another Mac, use its LAN endpoint.

# Connect Claude Desktop to this Mac
machboost launch claude-desktop

# Or choose a previously saved team host
machboost launch claude-desktop \
    --connection studio

# Restore your previous provider
machboost launch claude-desktop --restore

Also available in the app's Apps view. Requires Claude Desktop with third-party inference support and a compatible loaded model.

A few good questions

Before you
make it yours.

Ask us on GitHub

Is it faster for every prompt?

No. MachBoost's strongest measured gains come from reusable prefixes and repeated visual or repository context. Fresh, open-ended questions may run at native backend speed. The benchmark above compares the same model and weights, and names the workload.

Do I need Ollama or Python?

The Mac app includes its Python and MLX runtime, so neither a separate Python install nor Ollama is needed for native models. The standalone Python package and CLI require a compatible Python installation and the appropriate backend extras.

Can my teammates connect from another computer?

Yes. Enable authenticated LAN serving on the host, then connect with its network address and a scoped API key. MachBoost's app can use saved hosts; other compatible clients can use the host's API endpoint. The native desktop app requires Apple Silicon, while API clients can run on other platforms.

Does every Hugging Face model work?

No. Desktop inference supports compatible MLX text and MLX-VLM vision models. The app searches public Hugging Face repositories and checks architecture support before download. Reasoning, images, and tools depend on both the model and runtime support.

Does anything leave my Mac?

Local inference and chats stay on your Mac. Connecting to a host or paid provider sends the selected request there; an MCP connector follows its own service behavior. Model downloads and update checks contact their providers. MachBoost does not upload telemetry.

Why does macOS ask me to approve the app?

The free community build is ad-hoc signed, but not Apple-notarized. After the first blocked launch, open System Settings → Privacy & Security → Open Anyway. Verify the release checksum first. Supported installations receive updates through a signed Sparkle feed.

Your next local setup

Good models.
Great company. Your Mac.

Free to use. Open to build on.

Download MachBoost Explore the source

Latest community preview · Apple Silicon · macOS 14+

1
Download & drag
Open the DMG. Move MachBoost to Applications.
2
Approve the first launch
After macOS blocks it, use Privacy & Security → Open Anyway.
3
Make yourself at home
Download a compatible model or connect to a shared host.

Community preview, ad-hoc signed; not Apple-notarized. Release notes & checksum