# Shadowtype: On-Device AI Autocomplete for macOS

> A free open-source macOS app running a local LLM for inline ghost-text predictions in any text field, with no account and no network call for completions.

Published: 2026-09-03
URL: https://daniliants.com/insights/shadowtype-on-device-ai-autocomplete-for-macos/
Tags: on-device-ai, macos, local-llm, apple-silicon

---

## Summary

Shadowtype is a free, open-source macOS app that runs a local LLM on Apple Silicon to show inline "ghost text" predictions in any text field, replacing the need for cloud-based AI writing subscriptions. All inference happens on-device via llama.cpp and Metal, so there is no account, no network call for completions, and it keeps working with Wi-Fi off.

## Key Insight

- Positions itself directly against subscription AI writing tools by lifetime cost: Grammarly Premium runs $360 in year one and $1,800 over five years, and a comparable cloud rewriting tool runs $240 and $1,200. Shadowtype is a one-time download with all features unlocked forever.
- Works system-wide at the OS caret level (Mail, Slack, Notes, code editors, browser fields) rather than as a separate app window or copy-paste workflow, the differentiator most local-AI writing tools miss.
- Interaction model: `Tab` accepts the next predicted word, `Option-Tab` accepts the whole line, and `Option-Cmd-K` on selected text triggers an inline rewrite (formal, casual, shorten, fix grammar, summarize) that previews before committing.
- Ships a catalog of swappable GGUF models (Qwen3 Base for fill-in-the-middle continuation, Gemma 3, and MoE variants) and supports bring-your-own-GGUF, so the completion quality and speed tradeoff is user-tunable per machine.
- Exposes an OpenAI-compatible HTTP endpoint at `127.0.0.1` (`/v1/chat/completions`) plus a built-in MCP server, meaning it can double as a free local inference backend for scripts, editor plugins, or MCP clients, not just as a typing tool.
- Offers per-app and per-domain instruction overrides (casual tone in Slack, formal in Mail) and an on-device, encrypted, wipeable style profile that biases completions toward the user's own phrasing over time.
- The only outbound network traffic, ever, is an optional update check and the one-time model download. There is no telemetry and no churn email, because there is no account to churn from.