Local-first · Apple Silicon · Open source

MachBoost

Turn one powerful Mac into a private AI endpoint for your team, with resident models, repository-aware answers, and measured acceleration.

Download for macOS View source

Unsigned community preview · arm64 · macOS 14+

3.23× 7B repository questions
3.02× 3B repository workload
13.51× repeated-image median
Exact tokens in repository runs

Native control surface

Your Mac becomes the inference server.

Chat locally, keep models resident, attach persistent context, index a codebase, or serve the same runtime through familiar APIs.

MachBoost native macOS app showing a Qwen vision model, repository selector, persistent text and image context, and chat composer.
Native SwiftUI app with resident MLX text and vision models. No separate Python installation is required in the packaged build.

  • Repository intelligence

    Index tracked source once, retrieve focused evidence per question, and keep the reusable repository prefix stable.

  • Resident serving

    Keep multiple models warm behind one bounded, tenant-fair gateway with cancellation, employee limits, and resource metrics.

  • Text and vision

    Run MLX and MLX-VLM models with persistent files and images that remain attached until you remove them.

  • Drop-in APIs

    Connect existing tools through OpenAI- and Ollama-compatible endpoints or use the Python client directly.

Team Gateway

One Mac. Your team's coding agents.

Give each employee a scoped key and point Cline, Kilo Code, or another OpenAI-compatible client at the same private endpoint. Models stay resident while MachBoost enforces model access, concurrency, request rates, and fair queueing between employees.

Team Gateway guide
Team node Private network
Identity
Scoped keys
Admission
Tenant-fair
Protocols
OpenAI + Ollama
Tools
Parallel calls
export OPENAI_BASE_URL="http://TEAM-MAC:11435/v1"
export OPENAI_API_KEY="mbk_employee_key"

Govern access

Model allowlists, scoped keys, per-minute limits, and revocation from the native app.

Share repository context

Register a codebase once, retrieve bounded evidence, and keep reusable prefixes stable.

Inspect and evaluate

Choose metadata, redacted, full, or no traces, then evaluate selected requests locally.

Measured, not hand-waved

Speed where context repeats.

MachBoost accelerates exact reusable prefixes. It does not claim every first request or every unique prompt becomes faster.

Workload Model Native MachBoost Median Validation
Repository context Qwen2.5 3B 4-bit 3.144s 1.024s 3.02× 6/6 exact tokens
Repository context Qwen2.5 7B 4-bit 6.587s 1.998s 3.28× 6/6 exact tokens
Different repo questions Qwen2.5 7B 4-bit Five distinct questions 3.23× Exact tokens
Repeated-image matrix Six Qwen models 72 paired requests 13.51× 100% expected answers
First-view control Qwen3-VL No reusable prefix 0.99× No acceleration claim

Repository results use the same loaded MLX model, weights, prompts, and tokenization. Median prompt length was 10,405 tokens; 7,901 tokens were reused. The visual figure is the median of six model-level paired medians.

7B artifact Vision matrix Research paper

How it works

Reuse the expensive part safely.

  1. 01

    Load once

    The resident server keeps selected MLX models in unified memory.

  2. 02

    Build stable context

    Repository maps, documents, images, and chat prefixes are organized so exact reusable regions remain identifiable.

  3. 03

    Restore model state

    Compatible MLX caches skip repeated prefill work while preserving the same target-model generation path.

  4. 04

    Retrieve what changed

    Focused evidence and the new question are evaluated normally, with citations returned to the client.

Community preview

Download. Drag. Approve.

The preview DMG is free and ad-hoc signed. Apple has not notarized it, so macOS asks you to approve the app manually after download. Community updates open GitHub Releases and may require the same approval again.

  1. 1

    Download the DMG

    Apple Silicon and macOS 14 or newer are required.

  2. 2

    Drag MachBoost to Applications

    The bundled runtime means no Homebrew or Python setup.

  3. 3

    Attempt the first launch

    macOS will block the unidentified community build initially.

  4. 4

    Choose Open Anyway

    Go to System Settings → Privacy & Security, then approve MachBoost. Later launches of that installed version open normally.

Use your existing tools

One local endpoint.

The app and CLI speak through the same daemon. Point OpenAI-compatible clients at MachBoost or call the Ollama-style routes directly. Loopback works only on the host Mac; authenticated LAN mode displays a reachable address and token for teammates on the same private network.

OpenAI Python 
from openai import OpenAI

client = OpenAI(
    # Same Mac. For teammates, use the LAN URL shown in Server > Developer.
    base_url="http://127.0.0.1:11435/v1",
    api_key="machboost"
)

response = client.chat.completions.create(
    model="qwen2.5:3b",
    messages=[{"role": "user", "content": "Explain this repo"}]
)
Terminal 
machboost serve
machboost pull qwen2.5:3b
machboost run qwen2.5:3b