Introduction

Open-weight models on idle Apple Silicon, with Apple-attested hardware and a signed receipt.

Mifr is in public alpha. It runs open-weight models on idle Apple Silicon. Inference happens on the Mac: Apple-attested hardware, a logged build, a signed receipt bound to your request. The coordinator in the middle never holds a payload key.

You send a prompt to the hosted API and get the answer back from that machine.

If you already speak Chat Completions, that is a base URL and an API key. Compatibility is how you plug in, not the product.

Try it

Python

from openai import OpenAI

client = OpenAI(
    base_url="https://mifr-gateway-production.up.railway.app/v1",
    api_key="YOUR_MIFR_API_KEY",
)

completion = client.chat.completions.create(
    model="qwen3-4b",
    messages=[{"role": "user", "content": "Reply with the single word pong."}],
)

print(completion.choices[0].message.content)

Get a key

Sign in, mint a sk-mifr_ key, and send a prompt to an attested Mac.

See what ran

Every successful completion includes a signed mifr_receipt bound to your request.

Pick a Mac

List live providers at the hardware floor, or pin one by id on the next call.

What you can check

Apple-attested hardware and a signed receipt for the Mac that ran the request.

Origin

https://mifr-gateway-production.up.railway.app

SDKs want that origin plus /v1. The coordinator (wss://mifr-coord-…) and the website are different hosts. Mint keys at Settings; they start with sk-mifr_.

There is no first-party SDK and no /v1/responses. AI SDK 5’s createOpenAI()(id) hits /v1/responses, so use createOpenAICompatible (or openai.chat(id)).

Next steps

• Follow the quickstart to mint a key and get a completion.
• Client is the hosted API: auth, models, privacy, and the endpoint reference.
• Provider is for running a Mac on the network.