Free tool · no signup

AI agent permissions generator

Pick what your app holds. Get rules for what your customers' AI agents may see and do, and a policy file you can ship.

1Your product
2What a signed-in customer can see or do
See
Do
3How open to agents

Agents read with private fields hidden. Risky actions are closed to agents and wait for the owner otherwise.

4Start with

Nothing is blocked. Every decision is recorded with what would have happened.

See your 8 rules ↓

Your AI agent permissions

8 rules
  • Allow
  • Mask
  • Ask owner
  • Block
AI agent permissions for B2B SaaS or CRM, balanced
ResourceAI agentAI browserNot surePerson
Profile and contact detailsprofile.readAllow
Customer lists and searchcustomers.searchAllow
Documents and statementsdocuments.readAllow
Exports and CSV downloadsreport.exportAllow
API keys and tokensapi_keys.readAllow
Deleting recordsrecord.deleteAllow
Users, roles and invitesmembers.manageAllow
Account and security settingssettings.updateAllow

Tap an answer to change it.

AI agent permissions · B2B SaaS or CRM · Balanced · Watch first

Profile and contact details
  AI agent:     sees it with private fields hidden
  AI browser:   sees it with private fields hidden
  Not sure:     can read it
  Person:       can read it

Customer lists and search
  AI agent:     sees it with private fields hidden
  AI browser:   sees it with private fields hidden
  Not sure:     can read it
  Person:       can read it

Documents and statements
  AI agent:     is refused
  AI browser:   waits for the owner's passkey
  Not sure:     can read it
  Person:       can read it

Exports and CSV downloads
  AI agent:     is refused
  AI browser:   waits for the owner's passkey
  Not sure:     can read it
  Person:       can read it

API keys and tokens
  AI agent:     is refused
  AI browser:   is refused
  Not sure:     waits for the owner's passkey
  Person:       can read it

Deleting records
  AI agent:     is refused
  AI browser:   can ask; the owner approves each one with a passkey
  Not sure:     can ask; the owner approves each one with a passkey
  Person:       can do it

Users, roles and invites
  AI agent:     is refused
  AI browser:   can ask; the owner approves each one with a passkey
  Not sure:     can ask; the owner approves each one with a passkey
  Person:       can do it

Account and security settings
  AI agent:     is refused
  AI browser:   can ask; the owner approves each one with a passkey
  Not sure:     can ask; the owner approves each one with a passkey
  Person:       can do it

AI agent: proven agent activity in the session.
AI browser: an AI app's browser, no agent action seen yet.
Not sure: too little signal, never treated as a person.

Made with https://onehuman.ai/ai-agent-permissions-generator
Run these rules on your app →

Free for 30 days. Watch-only first, no card.

What AI agent permissions are.

Rules for what an AI agent may do with your customer's login. Each action gets one of four answers.

Allow

The request goes through as usual.

Mask

The page loads, private fields come back empty. The agent can still help.

Ask owner

The request waits until the account owner approves it with a passkey. An agent cannot approve itself.

Block

The agent is refused. A person takes the session back with a passkey.

One rule, four situations.

The same customer can be at the keyboard one minute and hand over to an agent the next. So each rule answers per situation.

AI agent

Agent tool traces in the page, a signed agent request, or clicks no hand could make.

AI browser

The page is open in an AI app's browser, but nothing shows an agent acting yet. The person's first real click unlocks it.

Not sure

Too little signal. Never treated as a person.

Person

Human clicks or a passkey. Always allowed.

Default AI agent permissions by data type.

The balanced defaults the generator starts from. A person is always allowed.

Default AI agent permissions by data type. A person is always allowed.
Data or actionAI agentAI browserNot sureWhy
Money dataBalances, statements, transactionsMaskMaskAllowRead-only money. Masking keeps the page usable.
Personal dataNames, contact details, ID documentsMaskMaskAllowRegulated data. Agents summarise it and pass it on.
Lists and searchCustomer lists, CRM searchMaskMaskAllowOne pass reads every record.
Medical or HR dataMedical records, salariesBlockMaskAllowThe most sensitive data an app holds.
Files and exportsDocuments, CSV exports, downloadsBlockAsk ownerAllowOne click hands over the whole dataset.
Keys and passwordsAPI keys, tokens, recovery codesBlockBlockAsk ownerMust never reach a model.
Moves moneyTransfers, orders, refunds, payout detailsBlockAsk ownerAsk ownerHard to reverse. A hidden instruction on a web page can trigger it.
Deletes or changes settingsDeletes, roles, invites, security settingsBlockAsk ownerAsk ownerHard to undo, and it changes who has access.

Questions

Something missing? Write to arif@onehuman.ai.

What are AI agent permissions?

Rules for what an AI agent may see and do when it acts with a customer's login in your app. Each action gets one answer per situation: allow, mask, ask the owner, or block.

Is this for AI agents my company builds?

No. It is for the agents your customers bring into your app with their own login. For agents you build, give each one its own identity and scopes in your authorization system.

How is this different from OAuth scopes?

Scopes cover apps a customer connects to your API. A browser agent never asks for a scope: it uses the customer's own browser, cookies and login. So the rules have to work inside the session, per request. How OneHuman does it →

Which AI agents do the rules cover?

Claude in Chrome, Claude desktop, ChatGPT agent, OpenAI Codex, Perplexity Comet, Playwright and Puppeteer. What each one does against a protected page →

Will my real customers notice?

Rarely. A person is always allowed, and their clicks are checked as they happen. In our tests 2 of 397 human clicks were read as a program. Start with watch first to see every decision before anything is enforced. How we measured →

What does "ask owner" mean?

The request is held. The account owner approves it with a passkey (Touch ID, Windows Hello), and the approval covers only that request: same action, same amount, same recipient.

Can I use the rules without OneHuman?

Yes. The plain-English rules work as a spec for any stack or a security review. The JSON file is OneHuman's policy format, and npx onehumanai init puts it to work on a Node.js server. Read the docs →

Is anything I pick here sent to you?

No. The generator runs in your browser. The share link keeps your choices after the #, which never reaches a server.

What does it cost to run the rules?

Nothing up to 100 customers a month whose AI agents use your app. You pay only for those customers. Pricing →

Run these rules on your app.Free for 30 days. Watch-only first, no card.