/ docsSandbox demo: registration to settlement#
Set spending boundaries in the CLI, run one governed request in Codex, and inspect the result.
v0.2.2 · 10:42 · Happy path · No real provider calls or charges
Command typing is silent and shown at 4× speed; long waits are shortened. Explanations and result review remain at normal speed.
The short version#
This companion script summarizes the recorded demo; it is not a verbatim transcript or an installation runbook.
1. Register the user and agent#
Hubu governs how AI agents spend money. In sandbox mode, the full Hubu, Gongbu, and Temporal stack runs locally with a simulated provider.
We check the version and service readiness, then register a user and an agent. Hubu verifies the registration fingerprints. The agent ID identifies the policy and budget subject; its account ID identifies the spending source.
2. Set policy and budget#
Policy defines what the agent may do. The recorded policy allows only Gongbu's local-fixture image scope, up to 10 simulated cents per request, and denies unmatched requests.
Budget defines how much it may spend. We give the agent one simulated dollar, expiring in 24 hours—a finite window, not a recurring allowance.
3. Ask Codex to do the work#
We switch to Codex, inspect the MCP tools, and select the same registered agent. CLI and MCP access the same governance state.
use the mock provider to generate an image of a blue dot with 1k resolution
Codex submits one request through hubu_submit_governed_execution. Hubu checks policy and budget and reserves funds before execution begins.
4. Follow execution and settlement#
Gongbu resolves the authorization, acquires an exclusive claim, and runs the mock provider through its Temporal workflow. It stores the artifact and reports the receipt to Hubu for settlement.
Codex follows that same operation to completion. The executor handles claim and settlement; the agent does not need to manage those steps itself.
5. Inspect the evidence#
The sandbox returns a fixed 1×1 PNG fixture, not a real 1024×1024 generated image. This tests the integration, not image quality.
Back in the CLI, we check the budget and ledger:
hubu budget list
hubu ledger list
The recorded cost is one simulated cent: USD 0.01 consumed, nothing frozen, and USD 0.99 remaining. The ledger records the settled cost; authorization records separately show requests and decisions.
Try it yourself#
People set the boundaries. Agents work within them. Results stay inspectable.
Follow the local stack quick start to prepare your own sandbox. This video begins with a running stack and focuses on the happy path. For real providers and approval decisions, continue with the local-stack demo. Failure handling will get a separate demo.