I need an AI writing assistant that works offline and does not send keystrokes to the cloud.

Yes. If your requirement is true local processing with no keystrokes sent to a cloud service, there are several good options right now.

Strong options

• Shadowtype — macOS: Probably the closest match if you want an AI keyboard/autocomplete experience. It runs an LLM locally on Apple Silicon, works offline, and explicitly says keystrokes never leave the device. It's open source and free.Shadowtype+1
• Stet — Windows/macOS: A local proofreading and rewriting assistant that works across applications. It uses a local llama.cpp model and is designed specifically around avoiding clipboard/cloud exposure.GitHub+1
• Polish — Windows/macOS/Linux: A system-wide writing assistant for fixing, rewriting, translating, summarizing, etc. Its Local mode runs on-device and offline.Polish+1
• Writtt — Windows/macOS/Linux: Better if you want a dedicated private writing environment rather than an assistant that operates inside every application. It integrates with Ollama, so the AI model itself runs locally.Writtt+1
• Draftside — Chrome: An offline writing editor using Chrome's on-device Gemini Nano. Its site says drafts stay in the browser and there are no backend AI calls.Draftside

If privacy is the absolute priority

I'd favor an architecture like:

Your keyboard → local writing assistant → local LLM (Ollama/llama.cpp) → your screen

rather than an ordinary AI writing service with a "don't store my data" privacy policy.

For example, Ollama explicitly states that when you run models locally, prompts and responses aren't transmitted to Ollama.Ollama

If you tell me Windows, Mac, or Linux (and roughly how much RAM/GPU you have), I can narrow this down to the most suitable 2–3 choices and explain exactly what data each one can access.

Yes. If your requirement is “the AI must run locally, and my text/keystrokes must not be sent to a cloud service,” there are good options.

What I’d look at

• lmstudio.ai + a local writing interface — LM Studio runs downloaded models entirely on your computer. Its documentation says that once the model is downloaded, chatting and document processing can work offline and entered text doesn't leave the device.LM Studio+1
• ollama.com + an editor such as Obsidian — Ollama runs the model locally, and there are Obsidian plugins that provide rewriting, continuation, grammar correction, summarization, etc. using only the local Ollama server.Ollama+1
• Obsidian + QVAC Local AI — particularly interesting if you want an integrated writing assistant. The plugin advertises local rewriting, grammar correction, expansion, semantic search, and other AI functions, with the model and processing kept on the machine.Obsidian Community
• Obsidian + Agentic Chat + Ollama — another local option; its documentation explicitly says that with Ollama, prompts, notes, and tool results stay on your device.Obsidian Community

One important distinction

If by “does not send keystrokes to the cloud” you mean you want something like Grammarly/Copilot-style assistance directly as you type anywhere on your computer, that's a different—and harder—requirement than a local chat/rewrite assistant.

For maximum privacy, I'd want the architecture to be:

Your keyboard → local application → local LLM → local result

rather than:

Your keyboard → cloud service → cloud LLM → result

And I'd disable optional web-search, telemetry, cloud-model, and remote-API features. Even applications that can operate locally may offer cloud features, so “supports offline” isn't quite the same as “guaranteed local-only.”

If you tell me Windows, macOS, or Linux and whether you want it to work in every application (Word, browser, email, etc.) or just in a writing app, I can narrow this down to the most suitable options.

Yes. If your requirement is “the AI must run locally, and my keystrokes must never be sent to a cloud service,” there are several good options right now.

Best options

• shadowtype.app — best for Mac autocomplete. It provides ghost-text completions directly at the cursor, runs the LLM on Apple Silicon, works offline, and explicitly says keystrokes never leave the device. It supports macOS 14+ on M1 or newer.Shadowtype
• brisktype.com — best for a Grammarly-like Mac experience. Local grammar correction, autocomplete and snippets, with no cloud processing and functionality even with Wi-Fi disabled.Brisktype
• typilot.com — best cross-platform option. Works on macOS, Windows and Linux and provides autocomplete, rewriting, dictation and writing polish across applications. Its AI runs locally through Ollama/Whisper; however, it does perform a periodic license check, so it isn't literally zero network traffic.Typilot
• draftside.ai — best free/open-source writing editor. It runs in the browser using Chrome's on-device Gemini Nano, with completions, rewriting, translation and other writing functions performed locally.Draftside
• openwriter.org — best for document-oriented writing. A local AI editor using llama.cpp, designed for rewriting, summarizing and grammar improvement without sending documents to external servers.OpenWriter

If you want the strongest privacy guarantee

I'd prioritize open-source software + a local model + the ability to physically disconnect from the internet.

For example, Shadowtype is particularly compelling if you're on a Mac: you can turn off Wi-Fi and it continues working, with its site stating that there is no inference network call.Shadowtype

If you tell me Windows, Mac, Linux, iPhone/iPad, or Android, I can narrow this down to the 2–3 best choices for your device, including which ones can work completely offline after installation.

To run an AI writing assistant 100% offline with zero data or keystrokes leaving your machine, you need a local LLM runner paired with an open-weights model. Because everything executes locally on your hardware, no internet connection is required, and your privacy is absolute.youtube.com+1

The best modern setup combines a backend engine to run the model and a user interface (UI) tailored for drafting or chatting.

1. The Core Engine: Ollama

• What it does: Ollama lets you run open-source large language models (like Llama 3, Mistral, or Phi-3) locally on Windows, macOS, or Linux.
• Why it fits: Once installed and the model is downloaded, you can completely disconnect your Wi-Fi. It runs entirely via your CPU/GPU with zero cloud telemetry.

2. The Writing & Chat Interfaces

Pair Ollama with a local UI depending on how you like to write:

• Jan: A polished, open-source desktop application that acts as an offline alternative to ChatGPT. It runs models locally out of the box and gives you a clean chat/writing interface without touching the cloud.youtube.com+1
• AnythingLLM: Excellent if your writing involves referencing local research, notes, or PDFs. It runs completely offline and lets you chat with or prompt-engineer text based on local document folders.youtube.com+1
• Obsidian + Local REST API / Ollama plugins: If you already write in markdown using Obsidian , you can install community plugins that hook directly into your local Ollama instance, allowing you to generate and rewrite text natively inside your offline notes.

Recommended Models for Writing

When setting up your engine, choose a model optimized for text generation:

• Llama 3 (8B) or Llama 3.1: Excellent prose quality, tone consistency, and instruction-following for general writing.
• Mistral (7B) or Phi-3: Lighter weight, highly responsive, and great for quick drafting or grammar fixes on standard laptops.

Would you prefer a setup guide for macOS or Windows , and do you know roughly how much RAM or VRAM (GPU memory) your computer has so we can pick the ideal model size?

Yes. If your requirement is “the AI must run locally, with no keystrokes or document text sent to a cloud AI provider,” I’d focus on local-LLM apps rather than ordinary cloud writing assistants.

Best options

• GPT4All — Probably the easiest starting point. It runs models directly on Windows, macOS, and Linux, requires no API calls, and can work completely offline after the model is downloaded.GPT4All
• AnythingLLM — My pick if you want a more polished writing/productivity assistant. Its desktop app runs on your computer, supports local models, and explicitly describes its document knowledge and AI processing as on-device.AnythingLLM
• Jan — Another good privacy-first desktop option if you want an open-source local assistant.
• LM Studio — Excellent for running local models, particularly if you want to choose among many models and have good hardware.

One important distinction

If by “writing assistant” you mean something like Grammarly that watches what you're typing inside Word, your browser, email, etc., the choices become much narrower. Many “AI writing assistants” send text to a server even if they advertise privacy.

For maximum assurance, I'd use a local model + a local writing interface, and ideally keep the computer disconnected from the internet while processing sensitive text. That makes the privacy boundary straightforward: the model literally cannot transmit your text.

If you tell me Windows/macOS/Linux + how much RAM you have (e.g. 16 GB/32 GB), I can recommend a specific setup and model that will work well for proofreading, rewriting, and autocomplete.