Delegate valuable work to AI.
Retain authority over the outcome.
Your browser does not have WebGL2 available. All the text, diagnostics and contact options are available in reading view.
For developers
Build agents that can do useful work. Add control at the moment an action becomes real, without replacing the models, tools or framework you already use.
One governed action
Wrap the tool call, not your entire application. Runtime evaluates the proposed action; your code remains responsible for execution.
Declare the agent, version, action and relevant context.
Respect the decision and map any effective-action changes.
Claim the authorised action and record its outcome.
Keep credentials server-side. Begin with your own sandbox tool. An error or pending decision is not approval.
Wrap one consequential tool call. Keep the rest of your code.
import { RuntimeClient, govern } from "@oorbital/sdk";import { executeRefund } from "./your-payments"; const runtime = new RuntimeClient({ credential: process.env.OORBITAL_RUNTIME_CREDENTIAL!}); const refund = govern({ client: runtime, proposal: (paymentId: string, amount: number) => ({ agentId: "agt_...", versionId: "agv_...", functionKey: "payments.refund", action: { functionKey: "payments.refund", target: { type: "payment", id: paymentId }, parameters: { amount, currency: "EUR" } }, context: { customerTier: "standard" }, environment: "sandbox" }), argumentsFromEffectiveAction: (action, [paymentId]) => [ paymentId, Number(action.parameters.amount) ]}, executeRefund); // Replace identifiers and your-payments with your own sandbox integration.
Runtime SDK example. Supply your own identifiers, credential and payment tool.
Integration
Your models, framework and tools stay yours. Add Oorbital where a proposed action becomes an execution.
refund(paymentId, amount)govern({ client, proposal, … }, tool)Only the authorised action.
The outcome is recorded.
A local illustration. No service is contacted.
Initialise RuntimeClient in your backend. Supply your agent and version identifiers. Keep secrets out of browser code.
Wrap the consequential tool call. Describe its target, parameters and context. Map the effective action back into tool arguments.
Start in sandbox. Test allow, review, refusal, changed actions and service failure before enabling production work.
Connect once at each relevant execution boundary, then manage the policy and evidence around it. Calls outside that boundary are not governed. A pending decision or unavailable service must not be treated as approval.
Control at execution
A useful boundary distinguishes ordinary work from a decision that needs you.
Approved information. An authorised recipient. Routine work can proceed.
Above the delegated limit. Hold the action for an authorised person.
An external recipient is outside the permitted scope. Do not share.
Interactive sandbox example.
Implementation, explained
Choose what matters to your integration. For the complete contracts and failure behaviour, start with the source documentation.
Open the quick start ↗Build something useful
Bring the agent you have, or start with the job you need it to do.
The door opened, but this destination isn’t here. Let’s get you somewhere useful.
Only known Oorbital destinations are suggested. No automatic redirects.