Painted lakeside terrace with AI conversations on two laptops, connected to a shared desktop Mac by a green cable.

A little less cloud. A lot more yours.

MachBoost.

Your AI, right at home.

A home for local models on your Mac.
Chat, code, and put your team's compute to work.

Download for Mac See it in action

Free community preview · Apple Silicon · macOS 14+

FIG. 01 / YOUR OWN LITTLE AI NETWORKModels run on your hardware.

THE COMPUTE BEHIND THE CONVERSATION

A local model, doing the work.

01 / YOUR REQUEST“Explain this API route.”From a laptop, app, or API client

02 / THE SHARED MACQwen2.5 3B · MLXModel weights stay resident

03 / STREAMED BACK

An API route receives a request, runs a handler, and returns a response.

Illustrative playback. No live model request.Requests go to the host. Answers come back to you.

A new home.
Familiar company. ChatGPT & Codex Claude preview OpenAI API Ollama API

THE EVERYDAY, WITH A LITTLE MORE COMPUTE

One place to make it happen.

A question, a code change, or a model for the whole team.
Start where you are. Keep your own workflow.

MachBoost / FieldnotesAUTO-PLAYING EXAMPLE

You

What should I know before the next release?

The release notes mention three changes: a new search endpoint, stricter input validation, and an updated client SDK.

Context: release-notes.md

Ask a follow-up...

Illustration of the workflowSample responses, not live inference

A PROPER HOME FOR YOUR MODELS

From “hey” to
“here's the diff.”

A native Mac app for conversations that become work.
Bring your files, your repository, and a model you choose.

MachBoost app with Muse Glimmer, repository tool calls, a code diff, and reasoning controls. Sample workspace from a UI test, captured September 6, 2026. — FIG. 02 / THE WORK, IN ONE WINDOWApp preview · sample workspace · September 2026

01 / CONVERSATION

A little more than chat.

Streaming answers, Markdown, reasoning controls, and saved conversations. Keep text files and images close at hand.

02 / WORKSPACE

Context meets action.

Give a compatible model repository tools. Follow its calls, choose edit permissions, and inspect the changes it makes.

03 / EXTENSIONS

Your tools are welcome.

Connect MCP servers and add custom instructions. Model capabilities are checked before you download.

Local chat history Bundled Python + MLX Compatible vision modelsDesktop guide

GOOD COMPUTE DESERVES GOOD COMPANY

One powerful Mac.
A whole team's AI.

Your teammates don't all need a workstation. Share a resident model from one Mac, then connect through MachBoost or a compatible coding tool.

Design's MacBookMachBoost app

Engineering's laptopCompatible coding tool

Your workstationOpenAI-compatible client

Your private network · authenticated requests

SHARED HOSTStudioResident models. A bounded request queue.

Ready to serve

Illustrative setup. Choose a preferred device or let the app rank compatible saved hosts.

Routing happens before generation. Each stream stays on its host; device memory isn't pooled. Repository tools run on the client. Paid API fallback is optional and explicitly configured.

Set up a shared host

LESS REPEATING. MORE GETTING SOMEWHERE.

Some work deserves
to be done once.

Reusable context can save the model from processing the same input again.
Here's a recorded example, with the conditions left in.

REPOSITORY-PREFIX REUSE

2.89×

median paired speedup

10 / 10 token-identical pairs

Same model. Same weights. Same prompts.

Lower wall time is better

Native MLX · cache off3.501 s

MachBoost · prefix reuse1.186 s

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

This is a historical repeated-context result, not a promise of faster generation everywhere. First questions and new context may get little or no speedup. A later Gemma/Muse audit found output differences on cached follow-ups; cache behavior depends on the model and runtime.

Recorded results Latency audit Run the benchmark

YOUR WORKFLOW, ALREADY INVITED

Familiar commands.
Your own compute.

Use the terminal, connect a coding client, or build on the same resident server that powers the app.

OpenAI Chat Completions & Responses
Anthropic Messages
Ollama-compatible endpoints

Runnable examples

# Install with the native MLX backend
python3 -m pip install --upgrade "machboost[mlx]"

# Download a model and start a conversation
machboost pull qwen2.5:3b
machboost run qwen2.5:3b

# Start the API server
machboost serve

The Mac app includes its runtime. This CLI install uses 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 key. Use its LAN endpoint for a shared Mac.

# Connect ChatGPT's Codex workspace
machboost launch chatgpt

# Launch Codex CLI with MachBoost
machboost launch codex

# Restore the previous desktop provider
machboost launch chatgpt --restore

Uses the Responses API. The CLI profile is separate from your usual provider configuration.

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

# Or use a saved shared host
machboost launch claude-desktop --connection studio

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

Also available in the app's Apps view. Long Claude Code tool loops remain a compatibility preview.

BEFORE YOU MAKE YOURSELF AT HOME

Good questions.
Straight answers.

Ask on GitHub

Is it faster for every prompt?

No. The strongest measured gains come from useful reusable context. A fresh question may run at native backend speed, and the server can add overhead. Benchmarks name the workload, model, cache state, and limitations.

Do I need Ollama or Python?

The Mac app includes its Python and MLX runtime. No separate Python installation or Ollama is needed for native models. The standalone CLI requires Python and the matching backend extras.

Can other computers use my Mac's models?

Yes. Enable authenticated LAN serving on the host, then connect with its address and a scoped key. MachBoost clients can use saved hosts; other tools can use a compatible API. Serving shares inference, not the machines' GPU memory.

Which models and file types work?

The desktop app supports compatible MLX text and MLX-VLM vision repositories. Reasoning, tool calls, and visual inputs depend on the model and backend. Attach plain text, source code, folders of text/code, and supported images. Check model capabilities before downloading.

Does anything leave my Mac?

Local inference and chats stay on your Mac. Shared hosts and paid-provider fallback receive requests you send to them. MCP connectors follow their service behavior. Downloads and update checks contact their providers. MachBoost does not upload telemetry.

Why does macOS block the first launch?

The community build is ad-hoc signed and isn't Apple-notarized. After the first blocked launch, use System Settings → Privacy & Security → Open Anyway. Verify the release checksum first. Supported installations use a signed Sparkle update feed.

MAKE A LITTLE ROOM FOR POSSIBILITY

Your next idea.
Your own machine.

Download a model. Open a conversation. See where it goes.

Download for Mac Explore the source

Latest community preview · Apple Silicon · macOS 14+

01
Download & drag
Open the DMG. Move MachBoost to Applications.
02
Approve the first launch
Privacy & Security → Open Anyway, after macOS blocks it.
03
Find your model
Download a compatible model or connect to a shared host.

Free community preview. Ad-hoc signed; not Apple-notarized.
Release notes & checksums