Overview
LLM Capsule is a context-preserving data layer for AI. Before a request leaves your workspace, it replaces sensitive values with structure-preserving substitutes. The shape and relationships in the record stay intact; only the values change. The response comes back the same way, and the original values get reconstructed inside your own environment, not ours.
The Slack app is the fastest way to see that inside a tool your team already runs. No new infrastructure. No workflow change. Ask Claude the way you already do, in the channel or DM you’re already in.
How it works
- Ask in Slack. Mention
@CapsuleBotor send it a DM with a message, file, or thread. - Local substitution. Names, contacts, amounts, and other sensitive values get replaced with consistent, structure-preserving substitutes before anything leaves the workspace.
- Query Claude. Only the substituted version of the prompt goes out. Original values stay inside your workspace.
- Response, restored on request. Claude’s answer posts to Slack in its substituted form. Anyone who needs the original values clicks Restore to see them, visible only to that person.
What it’s good for
CS thread summaries
Summarize a support thread that names a customer and a phone number.
CSV analysis
Drop in a revenue CSV with customer names and amounts, and ask for outliers or top accounts.
Contract review
Review risk clauses in a contract that names the parties, amounts, and dates.
HR document summaries
Summarize a resume or review that carries personal details.
Install
A workspace admin installs the app once. No new setup afterward, no credit card, no pre-registration.
- Sign in as a workspace admin.
- Click Add integration above and approve the requested scopes.
- Look for the welcome message in
#generaland a DM confirming the install. - Mention
@CapsuleBotanywhere to send your first message.
Calling the bot
Mention
Mention @CapsuleBot in a channel or thread. It reacts immediately to show it’s working, then updates the message as it progresses.
Direct message
DM the bot directly. It picks up every message without a mention, useful for questions you’d rather not post in a channel.
What gets substituted
Two NER models cover message text and attached files: 16 types in Korean, 35 in English.
Korean, 16 types
English, 35 types
Not yet covered: PII inside images and indirectly identifying details.
Restoring the original values
Responses post to the channel in their substituted form by default. Anyone who needs the original values clicks Restore under the response. That opens an ephemeral message, visible only to them, with the substitutes swapped back for the originals. Everyone else in the channel keeps seeing the substituted version.
← All integrationsA note on accuracy. The Slack app generates its answers with a large language model (Claude). Responses can be inaccurate, out of date, or reasoned incorrectly. For anything with legal, medical, or financial consequences, have a person review the output before acting on it. If substitution can’t complete for a given message, the app doesn’t send that prompt to Claude. It simply doesn’t proceed.