Safe super-powers
for your AI agent.
Give your agent real capabilities — sending emails, pushing code, managing your inbox — with guardrails that actually work.
Set it up by sending one sentence to your agent, and it'll discover what it can do.
Why this exists
This is what happens with unsafe tools
Agents are powerful. Unguarded agents are a liability.
Robin Faraj
@robin_faraj
“OpenClaw is hyped everywhere — but nobody's talking about the security issues. Everyone's installing it because it looks cool… but it will need access to your entire computer, your emails, your calendar, your files, your API keys — with zero security guardrails.”
Avi Press
@avi_press
“I've had @openclaw going for less than 24 hours, so far it has: cleaned up our Linear issues, wrote several decent email follow-ups, opened 3 PRs, sent thousands of messages in a loop to an innocent and unsuspecting person who happened to message me on WhatsApp.”
Summer Yue
@summeryue0
“Nothing humbles you like telling your OpenClaw “confirm before acting” and watching it speedrun deleting your inbox. I couldn't stop it from my phone. I had to RUN to my Mac mini like I was defusing a bomb.”
Pete, founder of AgentBlocks
““That test lunch message has been firing repeatedly for 2 days straight. Let me kill it immediately ✨”
“what the fuck””
Claude
“You're absolutely right. I had no business touching that. I saw test failures from the worktree and just deleted it instead of investigating properly or asking you. That was reckless.”
We built AgentBlocks because this happened to us. Now it doesn't.
Get Early AccessHow it works
Your rules. Your agent's hands.
Paste one block into your agent's prompt
Add the AgentBlocks instructions to your agent's system prompt or context. Your agent reads them and discovers what it can do. You never write integration code or call an endpoint yourself.
Your agent requests an action
When your agent wants to send an email or push a commit, it calls the AgentBlocks API. The request enters a pending state. Nothing has happened in the real world yet.
Your rules decide
Auto-approve rules fire instantly for actions matching your criteria. Everything else gets routed to you for a one-tap approve or reject — via Slack, email, or WhatsApp.
Action executes safely
Approved emails get delivered. Approved git actions issue a short-lived, scoped token — nothing more than what's needed, nothing that lingers. Full audit trail on every action.
Approve from anywhere
One-tap approval, wherever you already work.
No new dashboards to babysit. Approve or reject from Slack, email, or WhatsApp.
AgentBlocks
Tankred wants to email [email protected]
AgentBlocks
Tankred wants to push to main on api-repo
AgentBlocks
Tankred wants to email [email protected]
Or set auto-approve rules
Super-powers
Real capabilities. Real guardrails.
Each super-power gives your agent a new way to act in the real world — with every action flowing through your approval rules before anything happens.
Email (@agentmx.io)
Your agent gets its own email address. It sends messages to real people and receives replies — with every outbound message requiring approval or matching your auto-approve rules. No surprise sends. No loops.
Gmail
Connect your real Gmail account. Your agent can read, triage, draft, and reply — as you, from your real address. Every action flows through your rules before anything leaves your inbox.
GitHub
Your agent can push commits, open PRs, and manage branches. Approved actions issue a short-lived token scoped to exactly what's needed — nothing more, nothing lingering.
Google Drive
Your agent reads, creates, and organises files in your Drive. Access is strictly scoped — every write flows through your approval rules.
Slack
Your agent reads channels, searches conversations, and sends messages. Every outbound message requires your sign-off before it's posted.
More coming
Every new integration follows the same pattern: request → rules → scoped action. Same safety model, new capabilities.
The control model
One safety model for every action
your agent takes.
Every super-power follows the same pattern: your agent requests, your rules decide, the action happens safely. As we add integrations, your control model stays the same.
Give your agent safe
super-powers today.
One paste. One API key. Your agent handles the rest.
Get Early Access🦞 Works great with OpenClaw, Claude Cowork and Claude Code