MCP call_tool method returns CallToolResult object instead of content

[addypy]

Initial Checks

• I confirm that I'm using the latest version of Pydantic AI
• I confirm that I searched for my issue in https://github.com/pydantic/pydantic-ai/issues before opening this issue

Description

When using tools via MCP servers, the call_tool method returns the raw CallToolResult object instead of its content, causing inconsistency with how regular tools work:

Without MCP (regular Tool):

ToolReturnPart(tool_name='get_weather_forecast', content='Weather Forecast Response\nAction: search, Domain: weather-...')

With MCP Server:

ToolReturnPart(tool_name='get_weather_forecast', content=CallToolResult(meta=None, content=[TextContent(type='text', text='Weather Forecast Response\nAction: search, Domain: weather-...')]))

This inconsistency causes issues for those who expect the same structure regardless of how tools are implemented.

Proposed Solution

Modify the call_tool method in the MCPServer class to return result.content instead of the raw CallToolResult object. This would ensure consistent behavior between MCP tools and regular tools.

Affected Components

• pydantic_ai_slim/pydantic_ai/mcp.py

### Example Code

```Python

Python, Pydantic AI & LLM client version

python3.11
pydantic-ai==0.0.52
openai:gpt-4o,llama3.1:8b

[JohnUiterwyk]

I’m seeing this same behaviour.
The content of the ToolReturnPart is the python object from the mcp python sdk:

https://github.com/modelcontextprotocol/python-sdk/blob/main/src/mcp/types.py#L741

My take on this is that this is leaking framework implementation detail into the model interaction, the tool called by the model should return the result of calling that tool, not an unexpected object related to the frameworks orchestration the tool invocation.

A simple solution is to do a check on the mcp object and if it’s a single item and a CallToolResult and either a TextContent or ImageContent, the extract the string contents. ResourceContent could have similar logic or be left as is.

I’ll raise an PR with a proposed solution.

[timesler]

Here's a concrete example of this problem when using Anthropic models. This is the tool result that is actually provided to the LLM after a failed MCP tool call:

{
    "role": "user",
    "content": [
        {
            "tool_use_id": "toolu_vrtx_01Qe1CPFC3C2LKsVoaxDyQ2d",
            "type": "tool_result",
            "content": "{\"meta\":null,\"content\":[{\"type\":\"text\",\"text\":\"Error executing tool open_files: Error opening nonexistent.txt: File nonexistent.txt not found in workspace.\",\"annotations\":null}],\"isError\":true}",
            "is_error": false
        }
    ]
}

The problem is similar for successful tool calls, but the example above is additionally confusing due to the presence of both "is_error": false and "isError":true