Review MCP tools and side effects before agent use

Goal

Allow the agent to call only reviewed tools with permissions appropriate to the task.

Steps

1. Review name, description, and input_schema with MCP Client.list_tools(); obtain the complete list from a trusted server. [openai-mcp] [mcp-client]
2. Set tool_filter=create_static_tool_filter(allowed_tool_names=[...]) on MCPServerStdio or MCPServerStreamableHttp to expose only reviewed tools. [openai-mcp] [mcp-client]
3. Set require_approval="always" for write-capable tools, then attach the approved server to the agent under application policy. [openai-mcp] [mcp-client]
4. Add approved tools explicitly to agent configuration and disable others. [openai-mcp] [mcp-client]
5. Add a regression check that new remote tools do not automatically gain agent access. [openai-mcp] [mcp-client]

How to verify

Keep a per-tool review record; verify newly added tools are unavailable until approved.

Troubleshooting

If a tool appears unexpectedly, ensure configuration explicitly allowlists tools rather than accepting all discovered tools.