Desktop AI assistant with on-device masking

Shh! Don’t let the AI know.

Before a request reaches a cloud model, emails, card numbers, API keys and other details it recognizes are swapped for stand-ins on your computer — and restored when the answer returns. Local models work too.

macOS 12+ · Windows 11

// PRODUCT

A more private desktop AI assistant

Shhield AI runs on your computer, masks the details it recognizes before a request goes to the model, and restores them when the answer comes back.

Masked on-device first

Emails, card numbers, API keys and other details it recognizes are swapped for placeholders on your machine before your text goes to the model.

Stand-ins go to the cloud

Recognized details go out as placeholders such as [PA:EMAIL_1]. As the answer streams back, they are switched back to the real values on your screen.

Local models supported

Prefer to skip the cloud? Connect a local model so the model itself runs on this computer.

Masking is best-effort: anything the rules or models do not recognize is sent as written. Attached images and documents are not scanned, and tools and MCP servers receive the real values.

Local AI assistant →
// HOW IT WORKS

How it differs from ChatGPT or Claude in a browser

In a browser chat, what you type goes to the AI company exactly as written. Shhield AI is a desktop app that sits between you and the model you pick, and changes what leaves your computer. This is the path of one message, and where each step happens.

01

You write the message · on your computer

You type or paste text into the desktop app. Nothing has left your computer yet.

02

Details are recognized · on your computer

Built-in rules look for email addresses, card numbers, IP and MAC addresses, API keys, passwords and access tokens. Depending on your computer's region they also cover formats such as US Social Security numbers, UK National Insurance numbers, IBANs, or mainland China mobile and ID numbers. Words you add to your own list are matched too, and a language model you download can add names, organizations and places.

03

Each one becomes a placeholder · on your computer

An email address becomes [PA:EMAIL_1] and a key becomes [PA:SECRET_1]. The table that maps each placeholder back to its value stays in a local database; when your system keychain is available, the table is encrypted with a key kept there.

04

The masked request goes to your provider · over the internet

The app sends it straight to the provider you set up, with your own API key. It does not pass through a Shhield server. The provider receives the placeholders and every word that was not recognized, and its own data policy applies to what it receives.

05

The reply is restored · on your computer

As the answer streams back, each placeholder is replaced with the real value on your screen. The model wrote [PA:EMAIL_1]; you read the actual address.

06

Tools run with real values · on your computer or at the tool's service

When the assistant runs a tool, such as a shell command, a file edit or an MCP server, placeholders in the tool's input are filled in with the real values first, without asking you each time. What the tool returns is masked again before it goes back to the model, but a connected online service sees the real values.

With a local model through Ollama, LM Studio or the built-in engine, step 4 stays on your computer as well, and masking still runs. Other connections the app makes: at startup it downloads a public model catalog from models.dev, it checks for updates once a day, and it contacts shhield.ai to manage its license, sending an installation ID and your platform. None of these carry your conversations, and the app collects no usage analytics.

// EXAMPLE

What the model actually receives: a real example

Left is what you typed. Right is what the app's masking code produced from it, using the rules that are on in every region. Flip the switch to see the real values come back, as they do on your screen.

What you typed

Please draft a reply to [email protected]. Her test key sk-proj-Xr4Tq9LmZ2vB7nKw gets 401 errors when called from our server at 10.24.8.17, and the test charge on card 4242 4242 4242 4242 was declined. Explain both in plain words.

MASKED
Sent to the AI

Please draft a reply to [PA:EMAIL_1]. Her test key [PA:SECRET_1] gets 401 errors when called from our server at [PA:IP_ADDRESS_1], and the test charge on card [PA:BANK_CARD_1] was declined. Explain both in plain words.

Why these four were caught

The email address, the key starting with sk-proj-, the IPv4 address and the card number each match a rule that is on for everyone. Card numbers must also pass the Luhn check, so a 16-digit number that fails it is left alone. Each kind keeps its own count: a second email address in the conversation would become [PA:EMAIL_2], and the same address again would reuse [PA:EMAIL_1].

What this message would not hide

If it had said "Maya Lin from Lumenfold says invoice 2291 for $12,400 is wrong", the name, the company and the amount would go out unchanged under these rules. Names and organizations need a language model you download in Settings (the English one is still experimental), or you can add a client's name to your own word list. With your region set to the United States, a phone number such as (415) 555-0132 would go out as [PA:US_PHONE_1].

Keys and passwords in config files

The password rule takes the label along with the value. Paste the two lines db_host=10.0.0.5 and password = hunter22-not-real, and the model receives db_host=[PA:IP_ADDRESS_1] and [PA:SECRET_1].

Redact text online →
// MODELS

Cloud providers and local models it works with

Bring your own API key, or run a model on your own hardware. Every option below goes through the same masking step, local models included.

Ready in the first-run setup

Anthropic (Claude), OpenAI (GPT), Google (Gemini) and OpenRouter can be connected on the first screen by pasting an API key.

More built-in cloud providers

xAI (Grok), Mistral AI, Groq, Together AI, Fireworks AI, Perplexity, Cerebras, NVIDIA, Hugging Face, Vercel AI Gateway and Venice.ai, plus Azure OpenAI, Azure AI Foundry, GCP Vertex AI, Databricks, Snowflake and LiteLLM for company setups.

Models popular in China

DeepSeek, Alibaba (Qwen), Moonshot (Kimi), Zhipu AI, Z.AI, MiniMax and iFlytek Spark are in the provider list: pick one in Settings and paste your key. Doubao can be added from the Volcengine Ark template.

Local models

Connect to Ollama (it listens on port 11434) or LM Studio (port 1234) running on the same computer. The app can also download a GGUF model from Hugging Face and run it with its built-in llama.cpp engine, using the GPU through Metal on a Mac.

Any compatible endpoint

Add a custom provider for any server that speaks the OpenAI or Ollama API, such as a company gateway or a model you host yourself. Custom providers are masked the same way.

Not offered: agents that send their own requests

Coding agents that call a model on their own, such as Claude Code, Codex and Gemini CLI, were removed from the app because their traffic could not go through masking. The built-in assistant can still read files and run commands, and its requests are masked.

// DOWNLOAD

Choose the right version

Loading the latest release…

Verify downloads with the SHA-256 checksum.

// FAQ

Frequently asked questions

Does my original text get sent too?

Only what it recognizes is replaced; the rest of your message is sent as written. For everyone, it recognizes emails, card numbers, API keys and IP addresses out of the box. If your computer's region is set to mainland China, Hong Kong, Macau or Taiwan, it also recognizes mainland-format mobile and ID numbers, and forms of address such as a surname plus a title; full Chinese names rely on the Chinese language model, which is downloaded to your computer. Names in English, German and Japanese need that language's model, which you turn on in Settings. Images and PDFs you attach are not scanned, and tools the agent runs (including connected online services) receive the real values. Shhield AI never uploads your conversations, prompts or local files to our servers.

Does masking make answers worse?

Placeholders keep the type and a number: [PA:EMAIL_1] means "the first email address", [PA:BANK_CARD_1] "the first card". The same value always gets the same placeholder within a conversation, so the AI can still tell which one you mean; it just never sees the actual value.

Which models does it work with?

Mainstream cloud models with your own API key, plus local models.

Do I have to pay?

No. Every feature, including all privacy protection, is free. The optional Premium license (USD 29) currently adds a Premium identity in the desktop app and no other features. We plan Premium features such as more detailed statistics, with no committed scope or date; privacy protection itself will not become paid because of that.