Magika is entirely local — no external services, no network calls.

Here's what's happening under the hood:

• Google's Magika file type detection library is powered by

ort

, which is a Rust wrapper around Microsoft's ONNX Runtime. This means inference runs on a locally bundled ONNX model, entirely in-process. ort Docs

• Magika uses the high-performance ONNX Runtime for model inference and Tokio for asynchronous parallel processing — both of which operate purely on local CPU resources. Google Open Source

• The API works by creating a

Session

that can identify files from their path or from content read directly from memory — there's no remote endpoint involved anywhere in that flow. Lib.rs

The model weights are bundled into the crate itself (that's what the

sync.sh

/

gen

tooling in the repo is for — baking a new model into the library when one is released). When you call

identify_file

or

identify_content

, it feeds bytes into the local ONNX model and returns results without touching the network.

So you're safe to use it in air-gapped or privacy-sensitive environments. The only "AI" involved is a small neural network running locally via ONNX Runtime.