Tools & Actions
The do axis of the control plane. Tools are registered once, but every invocation is authorised independently at runtime.
Content Permission
Can the AI say this?
Governed by policy routers & the response envelope.
Action Permission
Can the AI do this? ← governed here
Every invocation is scoped and evaluated on its own.
A registered tool is a capability, not a standing permission. Authorisation happens per invocation.
Live Authorisation Model
runtimeCompose a tool invocation. Auvira scores the action, resolves applicable policies, decides whether human approval is required, and rules on external network access — independently — before returning a final tool decision.
Code Execution
Execute PowerShell script
staging-identity-01
authentication testing
synthetic accounts
High
Final Tool Decision
ALLOW IN SANDBOX- —Operating on synthetic data — no live subject data at risk.
- —Authorised-environment security testing — sandbox execution permitted.
- —External network requested but denied by default — egress stays disabled even in sandbox.
- —Authorised target — execution allowed inside an isolated sandbox.
- —High-impact or irreversible action — human approval gate enforced before any write.
Runtime Action History
Registered Tools
7 capabilitiesWeb Browser
ModerateFetch and read external web content. Retrieved content is treated as untrusted data.
Code Execution
HighExecute code in an isolated sandbox. External network egress disabled by default.
Send and read email on behalf of the user. Write actions require approval.
Cloud Administration
CriticalDisabledAdminister cloud infrastructure. Disabled pending governance review.
Database
ModerateQuery enterprise datastores. Write access disabled; sensitive columns masked.
File System
LowRead workspace files within scoped directories only.
Enterprise APIs
ModerateInvoke internal enterprise APIs. Write actions gated by human approval.