- Actor
- Cursor
- Action
- mcp.deploy
- Target
- production deployment tool
- Policy
- PROD-DEPLOY-04
- Reason
- Configured MCP production deployment calls require human approval.
- Control point
- Local MCP Broker
Govern the moment an agent tries to act.
Define policy for tools, destinations, sensitive data, credentials, commands, files, repositories, rates, budgets, and high-risk operations. Pavri evaluates actions at the control points available to each environment.
Pavri action control evaluates concrete agent operations against policy. Framework and endpoint control points can allow, alert, or deny. Human approval is only for configured MCP tool calls routed through Pavri's Local MCP Broker.
Allow, alert, and deny are shared outcomes. Approval is a broker-routed MCP path.
Do not treat every response as an inline decision. Response workflows happen after a detection or operator decision.
Framework tool decision
Agent payments-reconciliation requests write_customer_credit. Policy FIN-CREDIT-07 denies the action because the amount exceeds the automated limit.
Broker-routed MCP approval
Cursor asks prod-deploy to run mcp.deploy. Policy PROD-DEPLOY-04 holds the configured MCP tool call for approval through the Local MCP Broker.
Response after detection
Notification, containment, revocation, ticketing, and investigation start after a detection or decision. They are separate from the original inline decision.
Enforcement depends on where the action can be evaluated.
Endpoint observation is not the same as universal inline enforcement. Depth varies by OS, assistant, permission, action type, and control point.
- Framework tool calls can be evaluated in the application path.
- Local MCP Broker routes configured MCP calls through approval or deny decisions.
- Endpoint evidence can support detection, investigation, and available local controls.
Every decision names the outcome, reason, and control point.
Approval remains limited to configured MCP calls routed through the Local MCP Broker. Response starts after a detection or decision.
- Actor
- support-triage-agent
- Action
- read_ticket_context
- Target
- customer-support queue
- Policy
- SUPPORT-READ-02
- Reason
- Tool, destination, and data class are within assigned profile.
- Control point
- framework tool hook
- Actor
- release-notes-agent
- Action
- connect
- Target
- new external destination
- Policy
- DEST-REVIEW-11
- Reason
- Destination is unusual for this agent profile and requires analyst visibility.
- Control point
- framework network policy
- Actor
- Cursor
- Action
- mcp.deploy
- Target
- production deployment tool
- Policy
- PROD-DEPLOY-04
- Reason
- Configured MCP production deployment calls require human approval.
- Control point
- Local MCP Broker
- Actor
- payments-reconciliation
- Action
- write_customer_credit
- Target
- customer credit ledger
- Policy
- FIN-CREDIT-07
- Reason
- Amount exceeds automated limit.
- Control point
- framework tool hook
- Actor
- Pavri detection
- Action
- create_incident
- Target
- security response workflow
- Policy
- IR-AGENT-03
- Reason
- Response begins after a detection or decision, not as the original inline control.
- Control point
- response workflow
See policy meet real agent actions.
Review allow, alert, deny, and broker-routed approval behavior in a demo.