BETA Invite-only until Google verification

Give your AI agent only the emails it needs. Not your whole inbox.

You always know what you've shared, and you can see when each email was read.

Works with Gmail · No email stored

Our promise

“My assistant has the context it needs, and nothing more. I always know what I've shared, and I can see when it's been read.”

i.

Only the emails you pick.

Nothing is shared by default. Pick emails one at a time, or add a rule for a sender you trust Sender rules share email from one address you trust automatically. An email is shared only if Gmail confirms it really came from that sender, so a faked sender address doesn't get through.. Your agent can't read anything else, even if it asks.

ii.

Always know what you've shared.

The Shared tab lists every email your agents can read, and whether you shared it or a rule did.

iii.

See when it's been read.

The access history shows which agent was served which email, and when.

And you can stop it.

Revoke an email, and no agent can read it again through Permissible. Revoke an agent, and its next read fails.

Revoking stops future reads. It can't erase what your agent already read.

Connected agentTry it
ChatGPTActive
read “Your e-ticket” → served

What people use it for

Share a few emails.
Ask for real help.

The agent help you've been holding off on, without handing over your inbox.

Travel

“Build me an itinerary for the trip.”

Bills & receipts

“Where did my money go?”

Newsletters

“Give me a weekly digest.”

Job hunting

“Where does each application stand?”

Catching up

“What do I owe, and to whom?”

How it works

The connector is the gate.

Don't trust an agent to police itself. Your agent only ever receives what you've allowed.

  1. Connect Gmail

    Read-only. Permissible can't send, delete, or change anything.

  2. Pick emails

    Share individual emails, or add a rule for a sender you trust Sender rules share email from one address you trust automatically. An email is shared only if Gmail confirms it really came from that sender, so a faked sender address doesn't get through..

  3. Add it to your agent

    Add Permissible as a custom connector in ChatGPT, Grok, Muse, Cursor, and more.

Works with

Use the agents
you already use.

Any agent that supports custom MCP connectors can connect. We only mark an agent as working once we've tested it with a real shared email.

  • ChatGPT
  • Grok Bot
  • Muse
  • Cursor

Using something else? Tell us in your invite request. It decides which agents we test next.

Security

Why not just connect
Gmail directly?

9 of 10 models with inbox access read an emailed passcode and used it to raise their own wallet limit.

Read the study →

You decide. Not an AI.

Your agent only gets the emails you chose to share. No AI model reads your inbox or picks what to send, so no prompt can talk Permissible into returning an email you haven't shared.

We never act on your email.

Agents act on what they read: they draft replies and trigger workflows. Permissible doesn't. It never sends, replies, forwards, deletes, or does anything on your behalf, and our Gmail access is read-only, so it can't.

We keep IDs, not email.

Every email is fetched live from Gmail when it's opened. We never save the body, the subject line, or attachments, not even in your access history.

What we store

  • Your account email and name, to sign you in
  • Your Gmail access, encrypted
  • IDs of the emails you shared, or that a sender rule turned away
  • Sender addresses you set rules for
  • The agents you connected
  • Access history: which agent read which email, and when

What we don't

  • Email bodies, subject lines, or attachments
  • Your inbox list (it's fetched live each time)
  • Your search queries
We never sell your data, use it for advertising, or train models on it.

The honest limits.

A trust product should tell you what it can't do.

  1. You're trusting Permissible with read access so your agent doesn't need it. That's the trade, and it's why we keep so little.
  2. Revoking stops future reads through Permissible. It can't erase what your agent already read.
  3. A shared email can still contain malicious instructions. Permissible limits what an agent can learn from your inbox; it doesn't make an agent safe.
  4. The access history shows what we served, not what the agent did with it.

Why we built it

I wanted the agent help everyone else was getting, but not at the cost of giving an AI full access to my Gmail. So I built Permissible.

Built by someone who wouldn't hand over their inbox. Read the launch post →

Pricing

Free during the beta.

No payment during the beta.

BetaInvite-only
$0during the beta
  • Gmail
  • 5 connected agents
  • Manual sharing
  • Sender rules Sender rules share email from one address you trust automatically. An email is shared only if Gmail confirms it really came from that sender, so a faked sender address doesn't get through.
  • Access history
Request an invite →

FAQ

Questions,
answered plainly.

Something else on your mind? Ask it in your invite request.

If you need inbox access anyway, how is this safer than connecting my agent directly?
  • Agents act; we only gate. An AI assistant can draft replies, trigger workflows, or take actions based on emails it shouldn't have read. Permissible is a passive filter. It never takes action on your behalf.
  • Your inbox is the key to your other accounts. With full access, an agent could request a password reset somewhere else, read the code from your inbox, and take over that account. Through Permissible, it only gets the emails you shared, so it can't catch a reset email you didn't share.
  • You decide what's shared, not an AI. No AI model reads your inbox or picks what your agent gets. Only the emails you chose come back, so no prompt can talk Permissible into returning one you haven't shared.
  • We don't keep your email. No bodies, attachments, or inbox list, and we never train models on your data.
  • You're still trusting us. That's why our access is read-only, why we store so little, and why the code will be open source.
Do you read my email?
No. We open an email only when you, or an agent you authorized, asks for one you shared. Nothing from it is saved.
What happens when I revoke?
Future reads stop immediately. What the agent already read stays with the agent.
Which agents work?
Any agent that supports custom MCP connectors. ChatGPT, Grok Bot, Muse, and Cursor are confirmed. See the list.
Does it work with Outlook, iCloud, or other email? What about my calendar or documents?
Not yet. Permissible works with Gmail today. Other email providers, calendars, and documents are on our list. When you request an invite, tell us which ones you'd use.
Can I share with one agent but not another?
Not yet. Every connected agent can read your shared list, and you can revoke any one agent at any time.
Is it open source?
It will be, around the public launch.
Why does Google show an “unverified app” warning?
Verification of Gmail access takes weeks and includes a security audit. That process is underway, and in the meantime the beta is capped at 100 users.

Get the agent help you've been holding off on.

Request an invite. Tell us which agent you use and what you want help with. It decides which agents we test next.

We'll only email you about your invite.