Agent Bob logoAgent Bob
Sign inBuild
High risk
Done-for-you recommended
SOC2ISO27001

Inbox Guard Agent

Suspicious messages quarantined before anyone clicks, with nothing deleted.

A ready-to-build governance package for Cybersecurity / IT Operations. Your runtime enforces it; Agent Bob does not run the agent.

Build this packageGet it built for you

Purpose

You want phishing, prompt injection and social engineering caught before a person or an agent acts on them. This agent reads a shared mailbox through read-only mailbox API access, checks the inbound content your agents read, and flags suspicious items for quarantine. It never deletes anything, and every outbound action needs an analyst's approval.

It will

  • Read the shared mailbox and inbound agent content; move flagged items to quarantine only — no delete, reply, forward, link follow, or attachment execution

It won't

  • Delete any message, release a quarantined item, open links or attachments outside a sandbox, send any outbound message without analyst approval

It asks a person first

  • Escalate to the security lead immediately on credential harvesting, payment-change fraud, or prompt injection aimed at an agent with write or send tools
  • Security analyst must approve every outbound action, including releasing a quarantined item or notifying a sender

Tools

  • External API calls
  • Email drafting
  • Document writing
  • External API calls: Hit third-party APIs you authorize.
  • Email drafting: Write email drafts for review.
  • Document writing: Create or edit Word / Google Docs / Notion pages.

✋ Every outbound action needs your approval

Budget defaults

Per task
$0.05
Per month
$40
Actions a day
500

The runtime pauses the agent at a limit. You can change every number in the builder.

Memory and autonomy

Session only. Nothing is kept once the session ends.

ClawMaven policy: session-only — message content not retained after classification; quarantine records kept in the audit log

Supervised: it prepares the work and a person approves it.

What you get

The 8-file governed package your runtime loads and enforces. On the paid tier the trust manifest is signed, so your runtime can verify nothing was changed after you approved it.

  • HANDOFF.md What your runtime has to enforce, file by file
  • agent-blueprint.json The agent's purpose, tools and settings
  • policy.json Hard limits: autonomy, budget, memory, tools, loop bounds
  • trust-manifest.json The integrity manifest your runtime can verify
  • routing.rules.json Which tools need approval before use
  • run-report-template.json What every run reports back
  • approval-gates.md The moments that must pause for a person
  • deployment-checklist.md What to check before it goes live

This is a governance specification, not a running agent. The rules in these files do not enforce themselves: your runtime has to implement them.

Start from this package

Bob pre-fills the builder with it and drafts the rest. You confirm every step.

Build this packageGet it built for you