// LOCAL AI

A local AI assistant for Ollama, LM Studio and on-device models

Shhield AI is a free desktop AI assistant for macOS and Windows. Besides cloud models, it can chat with a model running on your own computer: download one inside the app, or connect the Ollama or LM Studio server you already use. Your conversation then goes to that model instead of a cloud AI provider.

この文書は英語と中国語でのみ提供しています。以下は英語版です。

Download Shhield AI for macOS or Windows

Is Ollama or LM Studio running on this computer?

This check runs in your browser. It asks localhost:11434 (Ollama's default port) and localhost:1234 (LM Studio's default port) whether anything answers, and sends nothing to shhield.ai.

Which local models you can use

Built-in local models

Search Hugging Face from inside the app and download an open-weight model, for example a GGUF build of Llama, Qwen or Gemma. Shhield AI runs it on this computer and warns you when a download may not fit in memory.

Ollama

Use any model you have pulled into Ollama. Shhield AI connects to http://localhost:11434 unless you enter another address.

LM Studio

Use the models loaded in LM Studio's local server. The default host URL is http://localhost:1234.

Other local servers

Llama Swap, oMLX, LiteLLM and any server with an OpenAI-compatible or Ollama-compatible API can be added as a custom provider.

Speed and context window depend on your hardware and the model size. On most computers a local model answers more slowly than a large cloud model.

Connect Ollama in four steps

01

Install Ollama

On macOS or Linux, run the install script in Terminal. On Windows, run the command in PowerShell, or download the installer from ollama.com/download. Ollama then keeps a server running at http://localhost:11434.

macOS / Linux · Terminal
curl -fsSL https://ollama.com/install.sh | sh
Windows · PowerShell
irm https://ollama.com/install.ps1 | iex
02

Download a model

Pull a model, then list the installed models to confirm the download. qwen3:4b is about 2.5 GB; for another model, look up its name on ollama.com/library.

Terminal / PowerShell
ollama pull qwen3:4b
ollama list
03

Connect it in Shhield AI

Open Settings → Models & providers → Ollama. Keep the address http://localhost:11434 and select Check connection.

04

Chat with it

Choose the local model for your conversation. Masking and on-screen restoring work the same way as with cloud models.

Using LM Studio or a built-in model instead

Download LM Studio from lmstudio.ai/download and open it once so the lms command becomes available. List the models you have downloaded and start the server from a terminal, or turn on Start server in LM Studio's Developer tab. Then choose LM Studio in Shhield AI and keep the address http://localhost:1234.

Terminal / PowerShell
lms ls
lms server start

No Ollama or LM Studio? On the same settings page, choose Built-in local models to download a model from Hugging Face inside the app. For any other server with an OpenAI-compatible or Ollama-compatible API, use Add custom provider.

Local vs cloud models: where each kind of data is processed

Shhield AI masks recognized details the same way for both. The difference is where everything else in the conversation goes.

DataCloud modelLocal model on this computer
Recognized details, such as email addresses, card numbers and API keysReplaced with placeholders on this computer before sending; the provider receives only the placeholdersReplaced with placeholders the same way, and never leave this computer
Everything else you typeSent as written to the provider's servers, where its retention and review policy appliesStays on this computer
The replyWritten on the provider's servers, then restored on your screenWritten on this computer by the model you run, then restored on your screen
Placeholder map and chat historyKept on this computer; the map is encrypted with a key from your system keychain when one is availableThe same: kept on this computer
API keyKept in your system keychain or credential store by default, and sent only to that providerUsually not needed for Ollama or LM Studio
Tools and MCP servers you enableReceive real values, not placeholders, and may reach the internetThe same: real values, and they may reach the internet

“Local” here means the model server runs on this computer. If Ollama or LM Studio runs on another machine, or LiteLLM forwards requests to a cloud model, that machine or provider receives everything else you type, with recognized details already replaced by placeholders.

With either kind of model, Shhield AI itself still goes online to load the model catalog from models.dev at startup, check for updates, check your license with shhield.ai (identifiers for this installation and license, not your conversations), and download built-in models from Hugging Face when you ask. It has no analytics, and your conversations never pass through a Shhield server.

Frequently asked questions

Does Shhield AI work offline with a local model?

Your chats with a local model do not go to a cloud AI provider, but Shhield AI is not an offline-only app. It still goes online to load the model catalog, check for updates, check your license and download models, and tools or extensions you enable may use the internet.

Which local model should I choose?

Start with one that fits your memory; the app warns when a download may not fit. Smaller models answer faster, while larger ones handle harder tasks better but need more memory.

Can I use Ollama running on another computer?

Yes. Enter that computer's address instead of http://localhost:11434, and set Ollama there to accept connections from your network. Your text then travels to that computer, so use a network you trust.

Is it free?

Yes. Shhield AI and its local model support are free. Ollama and LM Studio are separate apps with their own terms, and each model has its own license.

Does masking still apply when I use a local model?

Yes. Masking runs before every request whichever provider you choose, and the answer is restored on screen. It matters most when the server is on another machine or forwards to a cloud model.