One local server.
Three APIs your tools already speak.

Getting started

Start the server from the MLX Core app (it runs whenever a model is loaded) or from the CLI:

mlx-serve serve          # serves everything under ~/.mlx-serve/models, loads on demand
mlx-serve run gemma4     # or: download one model, serve it, and chat

The default port is 11234 (--port to change it). Models are addressed by name — the model field accepts a model's directory name, an org/repo id, or an Ollama-style short name with tag. A request naming a discovered-but-unloaded model loads it on the fly.

curl http://localhost:11234/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemma-4-e4b-it-8bit",
    "messages": [{"role": "user", "content": "Why is the sky blue?"}],
    "stream": true
  }'

Or with the OpenAI SDK — just point base_url at your Mac:

from openai import OpenAI
client = OpenAI(base_url="http://localhost:11234/v1", api_key="unused")
r = client.chat.completions.create(
    model="gemma-4-e4b-it-8bit",
    messages=[{"role": "user", "content": "Hello!"}])
print(r.choices[0].message.content)

OpenAI-compatible endpoints

Request extensions beyond OpenAI

All optional; every endpoint stays spec-compatible without them.

Sampling defaults resolve in this order: request body → server launch flags → the model's own generation_config.json → engine defaults. Clients that omit temperature/top_p/top_k get what the model's authors recommend.

Anthropic Messages API

POST /v1/messages implements the Anthropic Messages API — typed content blocks, tool_use / tool_result, thinking blocks, and the full SSE event lifecycle (message_start, content_block_delta, …). It's what lets Claude Code run against your local model:

export ANTHROPIC_BASE_URL=http://localhost:11234
export ANTHROPIC_API_KEY=unused
claude   # Claude Code now talks to your Mac

OpenAI Responses API

Ollama-compatible API

Tools built for Ollama — Raycast, Obsidian, Enchanted, Open WebUI, ollama-python/js — connect unchanged (the swap, in depth). Run on Ollama's port and nothing needs configuring at all: mlx-serve serve --port 11434.

Media generation

Each media endpoint accepts "stream": true for SSE progress events ending in a base64 complete payload; without it the response is the finished artifact. Load and unload media models like any other (below) — chat and media coexist in one process.

POST/v1/images/generations

FLUX.2-klein and Krea-2-Turbo (deep dive). Returns base64 PNG.

POST/v1/audio/speech

Qwen3-TTS text-to-speech and zero-shot voice cloning (deep dive). Returns WAV bytes.

POST/v1/audio/music-generations

ACE-Step text-to-music — describe a track, get 48 kHz stereo WAV (deep dive). New in v26.7.2.

POST/v1/video/generations

LTX-Video — text, image, or audio to video with synced sound (deep dive).

POST/v1/3d/generations

Hunyuan3D-2.1 — one photo in, a GLB mesh out; optional full-PBR texturing (deep dive). New in v26.7.2.

Response: {"created": …, "format": "glb", "data": "<base64 GLB>"} — standard glTF, opens anywhere.

Model management

Requests are queued per model and batch-decode together where the architecture allows; a media generation runs on the same GPU and briefly pauses chat decode — one machine, one memory budget, no surprises.

Metrics & auth

Both are opt-in and independent. Start the server with --metrics, --api-key, or neither — the app exposes the same two switches in Settings.

Recording happens once per request, never per token, so the metrics have no measurable effect on tokens/sec.

With --api-key <token> set, every request arriving from another machine must present the key — the OpenAI, Anthropic, and Ollama APIs, the index page, and the endpoints above. Any of these work:

curl http://mac.local:11234/v1/models -H "Authorization: Bearer <key>"
curl http://mac.local:11234/v1/models -H "x-api-key: <key>"
curl "http://mac.local:11234/v1/models?api_key=<key>"

Requests from the machine itself are trusted and need no key, so the app, your editor, and local scripts are unaffected — the key guards only what is reachable off the box. /health stays open either way. Because localhost is exempt, curl localhost will never show you a 401; test from another machine.