Without a card
Listing channels, direct messages and group chats, and resolving a message link someone pasted into the conversation. These calls are still audited.
Connector · Slack
Slack for your AI assistant, with nothing posted in your name unseen.
PrivacyFence connects an AI assistant to Slack with your own user token, so it sees the channels, direct messages and group chats you can see, and nothing more. There is no bot to invite. The assistant can find conversations straight away; it reads messages only after you have looked at them, and it sends nothing until you approve.
What the assistant can do
Find, read after review, send after approval.
Every Slack tool has a gate fixed in code. The Tools reference lists each one.
Reviewed before release
A channel's recent history, a thread's replies, and message search. You see exactly which messages would be released, then allow or deny.
Needs your approval
Sending a message to a channel or direct message, and starting a group chat. The card shows the conversation and the full text before anything is posted.
Personal data
Message text is checked before you see the card.
Before a reviewed read is shown to you, PrivacyFence scans the message text locally for likely personal data in English, German and Hungarian: IBANs, card numbers, ID and tax numbers, salary information and similar. Sender names and channel names are not scanned. Email addresses and phone numbers are deliberately not flagged, because nearly every signature has them.
A match sends the read to a card even if an always-allow rule covers the channel, highlights what was found, and asks for a second confirmation. The audit log records the category, never the matched text.
The privacy filter can also withhold whole kinds of Slack data from the assistant: message content, thread content, people's identities, and the channel, direct-message and group-chat lists. Each is set under Privacy Filter → Slack (in Settings on a desktop install, on the Admin page of an organization server).
Routine requests
Trust a channel, not the whole workspace.
An always-allow rule lets matching requests run without a card, and each one is still audited. For Slack, a rule can trust:
Named channels
Reading, or posting to, the channels and people you list. A search runs without a card only when every result is from one of them.
Public channels only
A search whose results all come from public channels.
Direct messages
One-to-one direct messages. This matches any one-to-one conversation, not only notes to yourself.
Group chats
Group direct messages, and starting one.
Likely personal data in a read still goes to a card. Every scope is in the scope catalogue.
Set it up
PrivacyFence's Slack app, or your own, then each person approves it.
- On an organization server, Slack works at once with PrivacyFence's built-in Slack app, with nothing to register; the sign-in redirect goes through the PrivacyFence relay. Slack limits apps that are not listed in its Marketplace, so the built-in app runs in limited mode: channel history is read through Slack search, and thread replies stop at 15. Slack sends no PKCE, so with the built-in app its one-time sign-in code passes through the relay without it.
- For full speed, an organization creates its workspace's own Slack app, with user token scopes only, and uploads its client ID and secret on the Admin page as a signed connector config; its sign-in does not touch the relay. A desktop install, which has no built-in apps, takes that app in its organization config bundle instead.
- Each person connects Slack from PrivacyFence and approves the app in the browser.
- The Slack tools appear to the assistant. The token stays with PrivacyFence and is never given to the AI client.
Other connectors
The same approval, for every service.
Each connector follows the same rule: reads are reviewed, writes wait for your approval. The others:
Next
See how a request is approved.
Follow a read and a write through PrivacyFence, card by card.