kua-agent
526b14ae00
Org-wide Claude Desktop plugin marketplace for MMD Group IT. Adds the mmd-it marketplace manifest and the mmd-workspace plugin (installationPreference: required) carrying four skills grounded in real MMD material: mmd-document-style, mmd-email-style, mmd-it-support, mmd-data-handling. Public repo, no secrets — MCP credentials are delivered separately via managedMcpServers.
3.6 KiB
name, description
| name | description |
|---|---|
| mmd-data-handling | Use before pasting, uploading, or discussing sensitive company or personal data in Claude Desktop as an MMD user, or when asked what is and isn't safe to share with the AI assistant. Triggers on "can I paste this", "is this safe to share with AI", "data handling", "confidential data in Claude", "what data can I give the assistant", or "AI policy". |
What can go into Claude Desktop at MMD
MMD Claude Desktop does not talk to Anthropic directly. Every request is routed through MMD's own AI gateway, running on MMD's own infrastructure, which sits in front of the actual model providers and presents them to Claude Desktop under MMD's own model labels. In practice that means the model actually doing the work behind a given "Claude" label in the app is not necessarily an Anthropic model — MMD's gateway maps some of the model tiers you see onto a different underlying provider's models, labelled so the experience is consistent, while still being routed and metered entirely through MMD's own gateway rather than a raw pass-through to one vendor.
The practical consequence: treat the assistant as an MMD-operated tool sitting in front of a third-party model backend, not as a private, Anthropic-only channel. Apply the same judgment you would to any other external processor of company data — the data leaves MMD's direct control and is processed by a model vendor outside MMD, even though the request and the keys are MMD's own.
Safe to use normally
- Ordinary work drafting: documents, emails, meeting notes, summarizing something you already have access to, general research and writing help.
- Company information you would already put in an ordinary internal email or document, at the sensitivity level appropriate to that channel.
Think before pasting
- Credentials, API keys, tokens, passwords. Never paste a live credential into a chat to "have the assistant use it" — every MMD automation pattern that needs a credential fetches it itself from a secured store; a credential typed into a prompt has left that store's protection and is now sitting in a model request.
- Bulk personal data about other people (employee records, customer PII, ID numbers, health or financial detail) beyond what the specific task actually needs. Give the assistant the minimum slice required, not a whole export "just in case it's useful."
- Anything under legal hold, an NDA that names AI processing restrictions, or a specific confidentiality marking that would already stop you from emailing it externally — if it wouldn't go in an external email, it should not go into a prompt either, since a prompt is also leaving MMD's boundary to a third-party model backend.
- Unreleased financial results, M&A, or similarly market-moving information — treat the assistant as an external party for the purposes of information-barrier rules that already apply to that category of data.
If you're not sure
Ask before pasting rather than after. Sensitivity judgment calls for a specific document or
dataset are exactly the kind of thing worth a quick check with MMD IT (see the
mmd-it-support skill) rather than guessing.
What NOT to do
- Don't assume "it's just Claude, it's fine" — the routing above means that assumption doesn't hold at MMD; the request goes through MMD's gateway to whichever backend model is actually serving that tier.
- Don't paste a secret/token to get something done faster — use the credential's own store.
- Don't upload a full customer or employee database when a filtered extract would answer the question.