Using Confidential AI in Excel, Word, and PowerPoint

Claude's add-ins for Excel, Word, and PowerPoint work on the file you have open, and its content goes along with every request. You keep using the add-ins, but the model behind them changes: instead of Anthropic's Claude models, open-weight models running inside confidential computing answer, so your cells, text, and slides stay encrypted in memory while they are processed.

Word, Excel, and PowerPoint icons
Introduction

The add-ins read more than you type

Illustration: Claude for Excel flags an unusual value in a forecast, with the message encrypted on your computer

The challenge

The file's content travels with every prompt

To answer a question about a sheet, a contract, or a deck, the add-in reads the content it needs and sends it along with your prompt. With a conventional provider, all of it is processed in plaintext on that provider's servers.

The solution

The same add-ins, a different backend

Desktop Proxy on your computer serves the Anthropic Messages API, so the add-ins can use it as their LLM gateway. It verifies the service through remote attestation and encrypts each request on your computer, and the model then runs in a hardware-isolated environment whose memory stays encrypted.

Benefits

Why use Privatemode with the Office add-ins?

Excel, Word, and PowerPoint

End-to-end confidentiality

Open-weight models, not Claude models

Free tier available

Video

Watch the setup, step by step

From the download to your first answer, for macOS and Windows. Pick your system and click a step to jump to it. The screens of the add-in and of the system are simplified illustrations.

Setup

Set it up in four steps

This guide covers the desktop versions of Excel, Word, and PowerPoint on macOS and Windows. You need Claude's add-in for each app you want to use, installed from Microsoft AppSource, and Desktop Proxy. The sign-in is the same in all three add-ins.

Privatemode portal sign-in screen

Get your API key

If you don't have a Privatemode API key yet, you can generate one for free here.

Desktop Proxy showing "Verified & serving" and its two local addresses

Install Desktop Proxy

Install Desktop Proxy for macOS or Windows and open it. In Settings, paste your API key and start the proxy. When the status page says "Verified & serving", the app has verified Privatemode and is ready.

The Integration settings of Desktop Proxy with the allowed origin https://pivot.claude.ai and the model prefix anthropic.

Allow the Office add-ins

In Settings → Integration, set Allowed origin to https://pivot.claude.ai and Model prefix to anthropic.. The add-ins run as web pages from that address, so the app has to allow it, and they only list models whose names start with the prefix. A page from the allowed origin can use the app to send prompts under your API key, but it never sees the key.

On macOS, keep HTTPS switched on under Settings → Local endpoint, which is the default, and allow the password prompt that trusts the app's local certificate. Office on macOS blocks plain http://localhost from add-ins. Windows needs nothing extra.

Connect the add-in

Open Claude's add-in in Excel, Word, or PowerPoint. On the sign-in screen, select Cloud provider or gateway, then Gateway. As the gateway URL, enter the app's local address from the box, without /v1, because the add-in adds /v1/messages itself. As the API token, enter any placeholder: the app uses the key from its settings.

Click Connect, then pick a model such as anthropic.glm-latest. Models ending in -latest follow Privatemode's model updates automatically. Do the same in the other Office apps you use.

FAQ

Frequently asked questions about Claude for Office with Privatemode

Check that the app's status page says "Verified & serving". Then check the gateway URL: it has no /v1 at the end, and it is http://localhost:8080 on Windows and https://localhost:8443 on macOS. On macOS, click Trust certificate on the status page if it says that browsers will refuse the HTTPS endpoint.

Integrations

View more

Explore other Privatemode integrations

Your agent remembers everything. Your model provider shouldn't see it.

Hermes Agent logo

Your gateway routes every team's prompts. Your model provider shouldn't read them.

LiteLLM logo

Talk to an expert

Have a questions about Privatemode? Let's talk!

Want to look for yourself?