Shadowtype finishes your sentences as inline ghost text in any app — and it does it without ever calling a server. There is no inference network request, period. Pull the Wi-Fi, board the flight, work air-gapped: it keeps predicting at full speed because the model runs 100% on your own Apple Silicon with llama.cpp and Metal.
Cloud autocomplete dies the moment your connection does — and it ships every keystroke to someone else’s server. Shadowtype runs the whole prediction loop on your Mac, so the network is never in the path between you and your next word.
Predictions are computed locally with a downloaded model. There is no API request to OpenAI, to us, or to anyone, for any completion. Turn off Wi-Fi and watch the ghost text keep flowing — same words, same instant speed.
Airplane Mode, a train tunnel, a hotel connection that drops every minute — none of it matters. Latency-free local inference means there’s no spinner, no “reconnecting,” no degraded suggestions when the bars disappear.
If your machine can’t talk to the internet by policy, Shadowtype still works. Nothing you type is transmitted because there’s nowhere for it to go — no telemetry backend, no analytics, no account. Set it up once, then disconnect forever.
Cloud completion bills per token and throttles you under load. Local inference is free to run as hard as your Mac allows — type all day, every day, with no usage meter on the model itself.
An OpenAI-compatible HTTP endpoint runs on 127.0.0.1 so your scripts, editors, and agents can call the same on-device model — no key, no cloud. A built-in MCP bridge plugs into Claude Code and other MCP clients, and the BYOM picker loads any GGUF.
Being honest about “offline” means naming every byte that could leave the Mac. Here is the complete list — and your keystrokes, prompts, and completions are on none of it.
Pick a GGUF model from the catalog (Qwen3 Base, Gemma 3, MoE variants — or bring your own) and it downloads once to your Mac. After that the file is local forever — re-launch on a plane and it loads straight from disk.
There is no license and no activation. Shadowtype is free and open source — it runs offline indefinitely with no recheck and no phone-home, from the moment you install it.
Shadowtype can check for a new version periodically so you get fixes and new models. It’s the only recurring outbound traffic, it carries none of your content, and you can simply ignore it.
Mail, Notes, Slack, your editor, a web form — Shadowtype predicts inline as you type. Tab accepts a word, ⌥Tab accepts the whole line. Offline changes none of it.
Select any text and press ⌥⌘K to rewrite it — shorter, clearer, a different tone — all computed locally. No connection required, your draft never leaves the machine.
Run a compact Qwen3 Base or Gemma 3 for speed, or an MoE variant for richer predictions — all downloaded once and executed on-device. Swap models anytime, or drop in your own GGUF; the inference path stays fully offline.
127.0.0.1 with no key required, so your terminal, editor plugins, and AI agents can call /v1/chat/completions against the same on-device model. A built-in MCP server plugs into Claude Code and other MCP clients, and the BYOM picker loads any GGUF you drop in. Everything is bound to localhost — nothing leaves your Mac.Download Shadowtype free, accept your first word with Tab, then pull the connection and watch it keep going.