Permission Model

📋 Planned

Documentation Under Construction

This page will explain Vanna's group-based permission model, including tool-level permissions, UI feature flags, and row-level security patterns.

Planned Content

  • ✓ Group-based access control overview
  • ✓ Tool-level permissions (access_groups property)
  • ✓ UI feature flags (showing/hiding tool details)
  • ✓ Row-level security patterns (filtering SQL results)
  • ✓ Permission checking flow
  • ✓ Example permission configurations
  • ✓ Audit logging for permission checks

Want to contribute or suggest improvements? Open an issue on GitHub

Three-Layer Permission Model

When complete, this will explain:

1. Tool-Level Permissions

class SensitiveTool(Tool):
    @property
    def access_groups(self) -> List[str]:
        return ["admin", "finance"]  # Only these groups can use this tool

2. UI Feature Flags

config = AgentConfig(
    ui_features={
        "tool_names": ["admin", "user"],       # Who can see tool names
        "tool_arguments": ["admin"],            # Who can see tool arguments
        "tool_results": ["admin", "developer"]  # Who can see full results
    }
)

3. Row-Level Security

class SecureSqlRunner(SqlRunner):
    async def run_sql(self, args: RunSqlToolArgs, context: ToolContext) -> pd.DataFrame:
        # Filter results based on user's department
        user_dept = context.user.metadata.get('department')
        filtered_query = f"SELECT * FROM ({args.query}) WHERE department = '{user_dept}'"
        return await self.execute(filtered_query)

Permission Flow

  1. User makes request → UserResolver extracts User
  2. User invokes tool → Check tool.access_groups against user.group_memberships
  3. Tool executes → Apply row-level filters if needed
  4. Results returned → UI features filtered based on user’s groups