You watched it refuse. Here is what getting one involves.
The agent in the demo is ours and every visitor shares it. What we provision is an agent with an identity of its own, its actions attributable to you, its scope set by you, its credential renewed and withdrawable. This page is the whole of what that is, what it does and does not do, and how it is arranged.
How to get one
There is no self-serve purchase, and here is exactly why
Two separate things stand behind that, and only one of them is about money. There is nothing to take money with yet: the company is incorporated, but a bank account and then payment-processor verification still have to follow, in that order, so there is no live key to charge a card with. Separately, and by design, provisioning is not self-serve: every account is created with no authority to spend, and the right to provision an agent is granted deliberately, by a person, after a conversation. Fixing the first would not open the second.
That is why there is no price on this page. What we offer instead is a short evaluation: tell us what you would point an agent at, and we will scope what a provisioned agent of your own would look like, and arrange it directly, not through a checkout.
What is and is not included
Stated here rather than in a footnote, because the boundary is the product. An agent that refuses by name is only worth buying if you know which refusals it can make.
Included
- One agent with an identity of its own. Not the shared demo credential, a key and a verifiable identity issued to you, attributable to you in the audit record, and withdrawable.
- Policy decided for that agent. The agent asks before it acts: the rules the substrate loads at startup are evaluated on every governed call, and a refusal names the rule that made it, the same path the demo just showed you.
- Deny by resource, action and subject. A rule can name what is being touched, what is being done to it, and which agent is asking.
- A credential that expires and renews itself. Identities are short-lived rather than permanent; the agent renews its own, which is what makes withdrawal mean anything.
- Google ADK or OpenAI Agents. Those are the two framework templates that exist today, alongside a plain base template.
Not included
- Multi-protocol coordination. Protocols are documented and carried on the spec; most have no wire behind them. Do not buy this for protocol interoperability.
- Delegation deeper than two levels. A root agent and the children it delegates to. A third level validates, is stored, renders back to you, and is ignored at runtime. Deeper trees are not supported.
- Governance of frameworks beyond those two. A provisionable, governed agent exists only for Google ADK and OpenAI Agents. Identity conformance is currently demonstrated for the LangChain signing adapter; the earlier ADK and OpenAI conformance runs are being redone after a fix. If you run something else, that is a conversation.
- An agent that does your work. You supply the model key and we hold none. What is governed is whether the call is permitted, not what the model generates.
- Self-serve provisioning. See above. The agent is granted, not bought, and the stack it runs on is woken on request rather than kept running.
What happens after you decide
It takes about three and a half minutes, and you will be shown it happening
Provisioning an agent is not instant and nothing about the design pretends it is. One provision through the public API, measured on 2026-08-21, returned in 217 seconds, the application log and the edge request log agree to the millisecond. That is one sample on a warm system for a one-agent spec, so treat it as the floor rather than the expected value.
The number that actually bounds it is the timeout envelope, not the measurement: the broker gives up at 420 seconds, deliberately before the platform would, so a failure is recorded as failed instead of sticking forever at "provisioning".
Both numbers are measured from a stack that is already up, and the stack is woken on request rather than kept running. So the 217 seconds is the provisioning call, not the wait: while the agent stack is hibernated the wake sequence comes first, and that part is arranged with you rather than measured in seconds. Added 2026-10-02, because a floor quoted without the thing that precedes it reads as the whole wait.
Stated because a wait nobody warned you about reads as a broken system. Three and a half minutes of silence after paying is the moment a first customer asks for a refund for something that worked.
- 0sThe request is accepted and your agent's row is written before any work starts, so its progress survives even if the connection does not.
- ~26sYour agent's identity is minted, the credential that makes its actions attributable to you.
- ~217sThe agent is deployed, governed, and answering. Measured, not estimated.
- 420sThe ceiling. Past this the attempt is recorded as failed rather than left pending, and you are told.
How to get one
Discuss an evaluation
Tell us what you would point an agent at, and whether it is Google ADK or OpenAI Agents. We will scope a short evaluation, say what it would involve, and arrange provisioning directly: there is no form that starts it on its own, because provisioning is granted deliberately, not bought.
Discuss an evaluation Book a call
Either button opens the same contact form. There is no mailing list and nothing else we send you.
And if you would rather not wait
The foundation underneath all of this is open source and installable now. You can mint an identity, sign a request with it and verify both against public keys alone, with no account and no key from us.