OpenAI Responses MCP

[cosbgn]

Description

OpenAI just announced support for MCP like this:

curl https://api.openai.com/v1/responses -i \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -d '{
    "model": "gpt-4.1",
    "tools": [
        {
            "type": "mcp",
            "server_label": "gitmcp",
            "server_url": "https://gitmcp.io/openai/tiktoken",
            "allowed_tools": ["search_tiktoken_documentation", "fetch_tiktoken_documentation"],
            "require_approval": "never"
        }
    ],
    "input": "how does tiktoken work?"
}'

I know also the ai-sdk has experimental_createMCPClient but as more API providers will have built in MCP I'm thinking that maybe it would be better to have a provider native way to use MCPs?

similar to openai.tools.webSearchPreview() we could also have openai.tools.mcp({options})

AI SDK Version

No response

[cosbgn]

Alternatively, we could also have a something like customTool() which would allow any json and pass it in the tool array:

like:

const result = await generateText({
  model: yourModel,
  tools: {
    my_mcp: customTool({ // this will not throw Unsupported tool type: mcp
            "type": "mcp",
            "server_label": "gitmcp",
            "server_url": "https://gitmcp.io/openai/tiktoken",
            "allowed_tools": ["search_tiktoken_documentation", "fetch_tiktoken_documentation"],
            "require_approval": "never"
        }),

  my_files: customTool({ // this will not throw Unsupported tool type: file_search
        type: "file_search",
        vector_store_ids: ["<vector_store_id>"]
  })

}
})

This way as providers add their custom tools we can always pass them directly without needing to add every one each time.