Loading…
Loading…
Short answers to what comes up before a first conversation. Anything more specific is usually covered on the integration, use-case or documentation pages.
A layer between an AI agent and the software your business runs on, which decides what the agent may reach, on whose behalf, and what needs a human decision first.
No. A chatbot answers questions. This is for agents that take actions in real systems — reading a customer record, updating an order, raising a ticket — where the interesting problem is authority rather than phrasing.
Yes. You configure OpenAI, Anthropic or Google Gemini per environment with your own provider key. The governance layer does not depend on which one you picked.
No. The whole premise is that the software you have stays where it is. You describe its API, review which operations an agent may call, and connect it.
No, and there is no field anywhere in the product that accepts one. A connection stores an opaque reference to a secret in your own vault. The runtime resolves it inside the execution layer, after the policy decision, so the value never reaches the model, the console or the audit trail.
Every run is bound to exactly one host tenant, and isolation is enforced with PostgreSQL row-level security rather than by application code remembering to filter. A query that loses its tenant predicate returns nothing rather than everything.
The tenant comes from the verified token, not from the conversation. No instruction in a document, a web page or a support ticket can widen it, because the conversation is not what sets it.
No. Run content is not used to train models.
The run pauses and shows the exact action with the exact arguments. Approving authorises that specific proposal, bound to a digest of it — not a category of future actions. Edit the proposal and the previous approval no longer applies.
Yes, and it is the recommended way to start. Publish only the read actions, run it for a few weeks, and decide which writes to publish from the evidence in the traces.
For reads, you get a wrong answer with a trace showing exactly what it read. For writes, it does not get that far without someone approving the specific action. The design assumes the model will sometimes be wrong and limits what that costs.
Kill switches stop new runs at the environment, application or deployment level without a deploy. Unpublishing an action makes it uncallable immediately, and revoking a connection fails closed rather than falling back to an older credential.
One plan at $99 per month, or ₹3,999 per month for customers billed in India, with the whole platform in it. Enterprise terms exist for single sign-on, private model routing and data residency.
No. End users are created automatically when people in your customers’ organizations first sign in, and they are not a billing dimension.
The assistant itself is a script tag and one endpoint on your backend. What takes longer is the decision work: which actions an agent should be allowed to call, and which need approval. That is the part worth not rushing.
Yes. The widget is a single script tag with no build step, and renders inside a shadow root so it cannot collide with your styles. Rails, Django, ASP.NET, WordPress and plain HTML all work.
The demo answers more in two minutes than this page does — it runs one agent across three systems and stops at the approval gate.