Permissions

The Permission Guard controls which tool calls require approval, which are auto-allowed, and which are blocked.

Permission Levels

Level	Behavior
Read	Auto-allowed (reading files, listing data)
Write	Requires approval (writing files, modifying data)
External	Requires approval (API calls, network requests)
Destructive	Requires approval with warning (deleting data)
Credential-sensitive	Always blocked unless explicitly allowed

Configuration

{
  "permissions": {
    "defaultPolicy": "confirm",
    "rules": [
      {
        "tool": "filesystem:read",
        "policy": "allow"
      },
      {
        "tool": "filesystem:write",
        "policy": "confirm"
      },
      {
        "tool": "github:*",
        "policy": "confirm"
      },
      {
        "tool": "shell:execute",
        "policy": "deny"
      }
    ]
  }
}

Policies

Policy	Effect
`allow`	Execute without asking
`confirm`	Pause and ask for user approval
`deny`	Block execution entirely

Approval Flow

When a tool call requires confirmation:

Jarvis explains what it wants to do
The approval panel shows the tool name, parameters, and risk level
You can Approve, Deny, or Approve Always (for this tool)
All decisions are logged in the audit trail

Audit Logs

Every tool execution is logged:

[2026-06-15 14:32:01] filesystem:write /workspace/README.md
  Risk: write
  Policy: confirm
  Decision: approved
  User: manual approval

View audit logs in Control Center → Permissions → Audit Log.

Next Steps

Permission Guard — deep dive into the safety model
Review Audit Logs — working with the audit trail
Security Model — architecture details