agentcloak-email-proxy_skill

Overview

This skill provides a secure email proxy for AI agents, letting them search, read, list, and draft emails without ever seeing raw credentials or unfiltered content. It enforces server-side credential isolation, PII redaction, HTML sanitization, blocklists, and prompt-injection detection. You can self-host or use the hosted beta depending on your trust and control needs.

How this skill works

AgentCloak sits between the agent and your mailbox and exposes a restricted API. Incoming email is processed by a four-stage pipeline: blocklist filtering, HTML sanitization, PII redaction, and prompt-injection detection. Credentials and OAuth tokens remain server-side; the agent only uses an API key. Drafts are created but never sent automatically, so a human review step is required before outbound messages go out.

When to use it

• Give an AI agent controlled read access to an inbox without exposing credentials
• Prevent leak of SSNs, credit cards, API keys, or other PII to an agent
• Reduce risk from malicious or cleverly formatted emails (prompt injection)
• Provide a sandboxed email workspace for research or automation workflows
• Self-host when you require full on-premise control over email data

Best practices

• Self-host if you cannot trust a third-party hosted service with your email credentials
• Create least-privilege API keys and rotate them periodically
• Enable and customize blocklists for financial and security senders to reduce false positives
• Review drafts manually before sending; never rely on automatic send by agents
• Test query translations (Gmail-style syntax) when using IMAP providers to ensure expected search results

Example use cases

• Automated inbox summarization for customer support agents while hiding customer PII
• Research agent that searches and extracts non-sensitive signals from newsletters and threads
• Draft reply generation for busy users where the human reviews and approves the reply
• Auditing or archiving workflows that need read-only access to mailboxes
• Secure testing of agent behaviors against adversarial or malformed emails