Support built-in MCPServerTool approval request

[Artui]

Description

Follow-up for this issue. MCP configuration provides a way for model to ask for approval from user to invoke an MCP tool. Currently it is not supported, needs to be done using Deferred tools

References

No response

[pydanticai-bot]

PydanticAI Github Bot Found 1 issues similar to this one:

1. "Native MCP support #3100" (95% similar)

[echarles]

@Artui Are you saying that it is possible today to implement user approval for MCP server combining MCPServerTool builtin tool and Deferred tools?

Any snippet would help...

... or it is not possible today and this issue is about making it possible?

[echarles]

@DouweM How related it this issue to #2889 ?

I need to add a feature where user is prompted before being allowed to use a tool, like Claude Desktop is doing. It is not clear to me from https://ai.pydantic.dev/mcp/overview what is not supported or not, and what would be the best way to move forward.

# From https://ai.pydantic.dev/mcp/overview
Pydantic AI supports [Model Context Protocol (MCP)](https://modelcontextprotocol.io/) in multiple ways:

[Agents](https://ai.pydantic.dev/agents/) can connect to MCP servers and use their tools using three different methods:
Pydantic AI can act as an MCP client and connect directly to local and remote MCP servers. [Learn more](https://ai.pydantic.dev/mcp/client/) about [MCPServer](https://ai.pydantic.dev/api/mcp/#pydantic_ai.mcp.MCPServer).
Pydantic AI can use the [FastMCP Client](https://gofastmcp.com/clients/client/) to connect to local and remote MCP servers, whether or not they're built using [FastMCP Server](https://gofastmcp.com/servers). [Learn more](https://ai.pydantic.dev/mcp/fastmcp-client/) about [FastMCPToolset](https://ai.pydantic.dev/api/toolsets/#pydantic_ai.toolsets.fastmcp.FastMCPToolset).
Some model providers can themselves connect to remote MCP servers using a "built-in tool". [Learn more](https://ai.pydantic.dev/builtin-tools/#mcp-server-tool) about [MCPServerTool](https://ai.pydantic.dev/api/builtin_tools/#pydantic_ai.builtin_tools.MCPServerTool).
Agents can be used within MCP servers. [Learn more](https://ai.pydantic.dev/mcp/server/)

[Artui]

@echarles hey - this has not been implemented yet

[DouweM]

How related it this issue to #2889?

@echarles Only insofar that they both relate to deferred/approval-required tools :)

This issue is about https://platform.openai.com/docs/guides/tools-connectors-mcp#approvals, which Pydantic AI could handle by essentially modeling that approval request as an approval-required tool call (raising the ApprovalRequired exception from https://ai.pydantic.dev/deferred-tools/) and then sending the result to the API as a mcp_approval_response.

If you're handling MCP requests agent-side rather than using the built-in tool, you could use https://ai.pydantic.dev/toolsets/#requiring-tool-approval to wrap the MCPServer.

In both cases, how to get the human's approval is up to you, but could be through a frontend protocol like AG-UI or Vercel AI. I don't think either currently have native support for this exact concept -- although you could likely get there with custom events or a special frontend tool.

[echarles]

although you could likely get there with custom events or a special frontend tool.

How to implement those custom events? Do I need #2382? (I am interested primarily in Vercel AI, AG-UI can also be useful).

[DouweM]

@echarles Are you using OpenAI MCPServerTool, or MCPServer with ApprovalRequiredToolset, or custom tools with requires_approval=True/raise ApprovalRequired?