AI agents can act.
Now they ask you first.
AllowedBy gives an AI agent a human's yes in one passkey tap, on the exact sentence the service wrote. Any service then checks who the agent is and who allowed it, offline, in a single request.
Agents are ready to act. The web has no way to say yes.
Assistants can now shop, sign up, pay invoices and change bookings. But nothing tells a service who the agent is, or which human allowed this action.
Agents borrow your keys
Today an agent works with your password, your API key or your card. The service can't tell it from you, and it can do everything you can.
Services fly blind
A good agent looks like a scraper. "Who allowed this payment?" has no answer an auditor would accept.
The agent writes its own question
"Confirm?" dialogs are written by the agent. An agent tricked by a web page can soften what it asks you to approve.
So services choose between handing over the keys and blocking agents. AllowedBy is the third option.
Ask, tap, verified.
The same three steps for a coffee order, a SaaS sign-up or a €50,000 payment run. Only the strength of the identity check changes.
The agent asks
It tries the action. Without a yes, the service answers with a link, a four-letter code and a sentence written by the service. The agent can't change that sentence.
You tap once
You open the link and check that the code matches the one in your chat. Then you approve with your passkey (Face ID, Windows Hello, a security key). You sign the service's sentence, not the agent's summary.
Any service verifies
The agent retries with two signed headers: its passport (who it is) and your grant (who allowed it). The service checks both offline. It never calls anyone and pays nothing.
# 1. no yes yet: the service answers with a link and a code HTTP/1.1 401 Unauthorized WWW-Authenticate: AllowedBy action="orders:place", level="1", approve="https://demo.upto.tech/a/01JB3K7Q…", code="K7QM" # 2. you tap · 3. the agent retries with proof; checked offline POST /orders Allowed-By: eyJ0eXAiOiJhYnktZ3JhbnQrand0Ii… AllowedBy-Proof: … HTTP/1.1 201 Created
Each service sets the level per action: reading a catalogue needs L0, paying an invoice L2, releasing a payment run L3 with two approvers.
One yes, three wins.
- Let your assistant do the errandsShopping, bookings, cancellations, sign-ups. It asks, you tap, it's done.
- Never hand over a password or a cardThe agent carries a yes that's limited to this action, this amount and this period.
- Budgets and an off switch"Up to €50 a week at this shop." See every agent and every grant in one wallet. Revoke any of them in one tap.
- Share only what you tickEach app knows you by a different id. Your name, email or address go only where you allow, and the wallet keeps the receipt.
Try it as a person
Open the shop, give its address to your AI assistant, and approve the order with your own passkey.
- Let good agents in, keep bots outKnow which operator runs each agent and which person stands behind it. Rate-limit by operator, not by CAPTCHA.
- One line to addWrap your app with the SDK. It publishes your rules, answers the 401 and checks every request before your code runs.
- Your words, your levelYou write the sentence the person signs and choose the level per route. You can also ask for verified data: name, email, address, "over 18".
- Evidence, not screenshotsEvery action carries a signed record of who approved it, at what level and on what wording. That settles a dispute.
For developers
from allowedby.sdk import AllowedBy, route aby = AllowedBy(origin=ORIGIN, node=NODE, org=ORG, routes=[ route("POST", "/orders", "orders:place", level=1)], templates={"orders:place": "Place this order"}) app = aby.wsgi(app) # every route checked
Your rules come out as OpenAPI 3.1, so agents and Claude (through MCP) read them on their own.
- See every agent acting for your peopleAn inventory of agents, the vendors they joined and what each one may do.
- Approval rules that follow the agent"Above €10,000, two people from finance with a verified identity." The rule holds in every SaaS, not only in your ERP.
- Offboard in secondsSync your directory (SCIM) and sign-in (SSO). When someone leaves, every grant and every agent key stops working everywhere.
- Answer the auditorSigned evidence packs: who approved what, at what level, and what was shared with which vendor. Built for the EU AI Act, GDPR and DORA.
Try the company console
Initech's console sets roles, budgets, vendor allowlists and co-approval rules. Every change takes a passkey tap.
The missing piece between identity and payments.
Identity providers know agents inside one company. Payment networks handle the money. Nobody carries a human's signed yes for any action, from one company to another.
| AllowedBy | Agent identity in an IdP | Agent payment mandates | Bot authentication | "Confirm?" in the agent | |
|---|---|---|---|---|---|
| Works across companies | ✓ | within one tenant | ✓ | ✓ | – |
| The person signs the service's own wording | ✓ | – | the basket | – | – |
| Who allowed it, checkable offline | ✓ | token only | ✓ | – | – |
| Identity strength per action (L0–L3, QES) | ✓ | login only | card holder | – | – |
| Actions that aren't payments | ✓ | ✓ | – | – | ✓ |
| Open, federated, any vendor or model | ✓ | – | per network | ✓ | – |
Built on standards you already trust
The consent layer for the agent economy.
Agents are moving from chatting to acting. Every action that matters will need proof that a person allowed it, at the right identity level and across company lines. AllowedBy is that proof, as an open standard with a commercial network behind it.
Why now
Assistants already shop, sign up and file expenses. Every EU member state must offer the EU Digital Identity Wallet by the end of 2026. The EU AI Act asks for human oversight of high-risk systems. The proof of a human's yes has no standard yet. The first network that services trust becomes the default.
Business model
Verifying stays free, so every service can accept AllowedBy. Revenue comes from the node network, which onboards agents, verifies people and sells company control and evidence. Finance approvals are priced per approver: €4–8 a month (est.). Merchants pay for outcomes, never for checks.
Agent-ready SaaS
"Continue with AllowedBy" on sign-up and login pages, enabled in minutes. Agents sign up to APIs and tools with a verified person behind them. Services pay per new customer. Supplier onboarding through Flowie comes first.
Fleet control for large companies
Their staff's agents are already on the network from step 1. Companies then buy the inventory, sharing policies, offboarding, approval rules and evidence, priced per employee.
Consumer actions
One ticket per person, cancellations, returns, age checks. Sold through fraud and bot vendors, who pay per signal.
Where it stands
Sixteen specification documents at version 0.2.9 (draft). A working reference implementation: home node, wallet, approval page, SDK, company console and five example sites, live on this domain. 193 end-to-end scenarios and 368 test vectors pass on every commit.
Moat
A two-sided network: services accept the grant because people carry one, and people enrol because services ask for it. A public transparency log, federated nodes and conformance profiles let the standard spread. The node network and the evidence it holds stay the business.
Market figures are published analyst ranges; analysts disagree by two to four times. Figures marked est. are our own assumptions.
Try it now, with your own passkey.
Five example sites accept agents, and one home node holds your wallet. Everything is live and real, except that identity checks for L2 and L3 are simulated.
On shop.demo.upto.tech, buy coffee beans and a mug, delivered to me. You'll get a link to approve with your passkey.Let agents do the work.
Keep the yes with people.
AllowedBy is an open standard for human approval of AI agents, and a network that makes it work.