live · no reviewed actions yet

Agents that can't go rogue.

PonyLayer sits between your AI agent and your inbox. It lets safe actions run, pauses risky ones for your approval, and keeps a full record of everything — so your agent stays useful without ever going off the rails.

No mailbox credentials sharedOne-click approvalsFull audit trail
Your agent AI
◼︎
Wants to reply to a customer
action: send_email
to: "all_customers"
body: "shipments delayed…"
PonyLayer policy
Checks every action
auth verified
no prompt injection detected
hold: bulk send needs approval
Your email gmail
Stays under your control
The difference

Same agent. Same request. Very different day.

"Agent — reply to every customer about the shipping delay." Here's how that plays out.

Without PonyLayer
The agent has the keys. And sometimes loses them.

Nothing stands between a misread prompt and 10,000 outgoing emails.

Mailbox credentials are sitting inside your agent's runtime.

When something breaks, you're piecing logs together to find out what happened.

A cleverly worded customer email can steer the agent into sending what it shouldn't.

With PonyLayer
You kept the keys. The agent kept its job.

Every action is checked against policy before it runs — out of the box.

Risky sends pause for a one-tap approval. Safe reads keep moving.

Every decision is timestamped, explained, and replayable.

Inbound content is screened before your agent reads it — no more injected instructions.

How it works

Three steps. Fifteen minutes.

i
Connect email

Authorize PonyLayer once. Choose which actions can run on their own, and which pause for you.

providergmail
autoread, label, draft
asksend, delete, forward
ii
Point your agent at us

MCP server, skill, or SDK — however your agent likes to talk. We speak the dialect.

mcp://ponylayer.com/email
statusconnected
iii
Approve on your phone

Sensitive sends ping you for a one-tap approval. Everything else just works in the background.

pending1 action
approved42 today
Features

Built for teams moving fast with agents.

Policy gates by action

Let reads through. Pause sends, deletes, and forwards. Dial in anything in between.

Human approvals

A tap on your phone, a click in Slack. No dashboards to live inside.

Prompt injection filter

Suspicious inbound content gets screened before your agent ever sees it.

Audit log by default

Every request, check, decision, and send — timestamped, traceable, exportable.

Zero credential sharing

Your agent never touches mailbox passwords or tokens. They stay with us, encrypted.

More connectors, same rules

Email's live now. Calendar, docs, and CRM are on the way — under the same safety model.

FAQ

Fast answers.

How does PonyLayer keep my mailbox safe from the agent?

Your agent talks to PonyLayer, not to your inbox directly. We hold the credentials and check every single action against your policy. Only the ones you allowed — or the ones you explicitly approve — actually run.

Do I have to swap out my agent framework?

No. We sit in front of your tools, not behind your model. Any agent that can call an MCP server, run a skill, or hit an SDK can use PonyLayer. Keep what you're already using.

Will this slow down everything the agent does?

Only the actions you flag as sensitive pause for approval. The rest fly through — usually in under 50ms of added latency.

What counts as a "sensitive" action?

You decide. Most teams let reads and drafts run automatically, and ask for a tap before anything leaves the building — sends, deletes, replies outside the org.

What about prompt injection?

Inbound email content is scanned for suspicious patterns before your agent reads it. Known injection tricks get flagged, and your policy decides what to do next.

Is this just for email?

Email is the first connector we shipped. Calendar, docs, and CRM are next. Same product, same policy engine, just more things your agent can safely touch.

Can we delete our data?

Yes — any time, with one click. We only store what's needed to enforce policy and keep the audit log.

Let your agent run. Keep the reins.