PaglaAI

PRODUCT

PaglaMLX

LLM orchestration, optimized for your Mac. One local API for MLX models and cloud providers, managed securely from the menu bar.

Read the docs View on GitHub

Overview

PaglaMLX is an orchestration layer that speaks both OpenAI and Anthropic natively, dropping straight into any existing AI workflow. Local MLX models run as isolated servers on their own ports; the gateway routes by name, falls back to cloud providers by prefix, and translates protocols in both directions — streaming and tool calls included.

Why it exists

Testing a local model used to mean spinning up isolated servers and managing ports by hand. Switching between local and cloud meant rewiring clients. PaglaMLX removes that ceremony.

  • Manually starting and tracking a server per model.
  • Reconfiguring every client to reach a different port or protocol.
  • No clean fallback when a local model is unavailable.
  • No single place to see what is loaded and what is running.

Solution

One gateway on 127.0.0.1:2525/v1. Every model is a directory; every provider is a prefix. Clients point at one address and let the router decide.

One API, two protocols

OpenAI /v1/* and Anthropic /v1/messages, translated bidirectionally.

Auto-routing

model=auto picks the best local model; gpt-*, claude-*, gemini-* and friends route to cloud.

Menu-bar control

Load, unload, and watch models without touching a terminal.

One-click integrations

VS Code, Copilot, Kilo Code, OpenCode, Cline, Continue.dev, Claude Code, Codex CLI and more.

Key features

Smart gateway

OpenAI-native and Anthropic-native endpoints on one local port.

Model orchestration

One-click load and unload; each model runs as an isolated mlx_lm.server.

Live route table

Routes update as models start and stop; multiple models run concurrently.

Cloud fallback

Prefix routing to OpenAI, Anthropic, Gemini, OpenRouter, Groq, and Together.

Free Router

A free fallback path through OpenRouter when nothing else is configured.

Secure by default

Local-only binding and a key from Settings, readable by your tools only.

Architecture

  • SwiftUI App (menu bar)

    Loads models, exposes settings, watches the event log.

  • RoutingGateway (Python, :2525 /v1/*)

    Translates OpenAI ↔ Anthropic, routes by name and prefix, streams.

  • Local Router

    One mlx_lm process per model, each on its own port (50xx).

  • Cloud Router

    OpenAI, Anthropic, Gemini, Groq, Together by prefix.

  • Free Router

    OpenRouter fallback when no cloud key is present.

Documentation

bash 
$ git clone https://github.com/paglaai/PaglaMLX.git
$ cd PaglaMLX
$ ./build_dmg.sh   # release build + .app + .dmg

Getting started

Set your models directory, check Python, load a model, and connect any OpenAI or Anthropic-compatible tool.

Start the guide

Downloads

PaglaMLX macOS app
.dmg · macOS 14+ · Apple Silicon
Download
Source code
.zip · MIT · Python + Swift
Source
Release notes
CHANGELOG · v1.5.0 current
Notes

Roadmap

  • Shipped

    v1.5.0 — Rebranding release

    Project renamed to PaglaMLX, docs and API reference published, release pipeline live.

  • Now

    Gateway hardening

    Event-log watcher, deterministic replay tests, and deeper provider coverage.

  • Next

    Model marketplace

    Browse, download, and pin models from the menu bar.

FAQ

Download the latest .dmg from Releases and drag it to Applications. From source, clone, run ./build_dmg.sh, and you're done.

Run it locally tonight

Open the quick start or clone the repo. Your first local model is minutes away.

Get started View on GitHub