AI connector
Use your own Claude or ChatGPT with client files.
Upload a client document to Safe-Desk. Your AI reads it with placeholders such as [SSN_1] where the client details were, drafts with them, and hands the draft back. Safe-Desk fills in the real values, and you open the finished document on Safe-Desk. The AI never sees a value your firm's rules swap out, on the way in or out. Anything your firm set to Flag, such as names or addresses, is sent as written and reported to the owner.
If your firm uses strict mode, files with identifiers such as SSNs are refused, and names and contact details are removed and not filled back in.
Before you start
You need a Safe-Desk account at a firm with an active subscription, and at least one file under Files in Safe-Desk. The connector address is the same for everyone:
https://desk.sophiaxt.com/mcpIn Claude
- Open Settings, then Connectors.
- Choose Add custom connector. Give it a name such as Safe-Desk, and paste the address above.
- Claude opens a Safe-Desk page. Sign in if asked, check that the page names Claude and your account, and choose Allow.
- In a chat, turn the Safe-Desk connector on from the tools menu and ask, for example: "List my Safe-Desk files, read the intake notes, and draft the engagement letter."
On Team and Enterprise plans an owner may need to add the connector for the organization first. Claude Desktop uses the same connectors as claude.ai.
In ChatGPT
- Open Settings, then Connectors. Under Advanced, turn on developer mode.
- Choose Create. Name it Safe-Desk, paste the address above, and choose OAuth for authentication.
- ChatGPT opens a Safe-Desk page. Sign in if asked, check the page names ChatGPT and your account, and choose Allow.
- In a chat, add the Safe-Desk connector from the tools menu and ask it to work on your files.
Developer mode and custom connectors depend on your ChatGPT plan and your workspace's settings.
Both apps change these menus from time to time. If a step looks different from what you see, email contactus@sophiaxt.com and we'll walk you through it.
What the connector can do
| Tool | What it does |
|---|---|
list_files | Lists your own files: title, when uploaded, and what kinds of client data were replaced. Titles show placeholders too. |
read_file | Gives the AI one file's text with placeholders only, and asks it to keep them exactly as written. |
finish_document | Takes the AI's draft and the files it used. Safe-Desk fills in the real values from those files and saves the result for you. The AI gets back a link and a count, never a value. A placeholder with no saved value is left as written and named, so you can fill it in by hand. |
Every tool works only on the signed-in person's own files. A colleague's files, another firm's files and your firm's settings are out of reach, whatever a document or a prompt asks for.
What it doesn't do
A connector runs after the AI has already read your chat, so it can't protect what you type or paste into Claude or ChatGPT. The Safe-Desk Chrome extension does that; Settings in Safe-Desk has it. Put client documents through Files instead of pasting them into the chat.
What is stored, and for how long
When you upload, Safe-Desk runs your firm's rules once. Passwords, keys and privileged material are refused, and nothing is kept. Otherwise it keeps the protected text and the real values, both encrypted and tied to your account, for 7 days, or until you delete the file. The original file is not kept. Finished documents are also encrypted and deleted after 7 days. Your firm's owner sees in the log that a file was uploaded, read or finished, and which kinds of data it held, never the values.
Reading and finishing count toward your firm's monthly AI allowance at about four characters a token.
Disconnecting
Settings in Safe-Desk lists every AI app you've connected, with Revoke. Removing the connector in Claude or ChatGPT, changing your password, or being switched off by your firm's owner also ends the connection.
For developers
The connector is a remote MCP server over Streamable HTTP (JSON responses) at /mcp, protocol versions 2025-06-18 and 2025-03-26. Sign-in is OAuth 2.1 with PKCE (S256) and dynamic client registration; discovery starts from the 401 on /mcp and /.well-known/oauth-protected-resource. Access tokens last an hour and refresh tokens 30 days, rotated on use. Redirect addresses must be on claude.ai, claude.com, chatgpt.com, or loopback for desktop apps; ask us if you need another.