MCPTools now discovers every page of tools from a ClientSession

MCP servers can split their tool list into pages and hand back a next_cursor for the next one, and Agno's MCPTools now follows that cursor on a ClientSession you pass in. Before, any tool on a later page never reached your agent, and a tool named in include_tools that lived on page two failed validation.

This run uses an in-process FastMCP server that lists one tool per page:

from fastmcp import Client, FastMCP
from agno.tools.mcp import MCPTools
 
server = FastMCP("docs", list_page_size=1)
 
@server.tool
def search_docs(query: str) -> str:
    return "results"
 
@server.tool
def read_page(path: str) -> str:
    return "page"
 
async with Client(server) as client:
    tools = MCPTools(session=client.session)
    await tools.build_tools()
    print(list(tools.functions))

Before: ['search_docs']
        include_tools=['read_page'] -> ValueError: Included tool(s) not present in the toolkit: read_page
Now:    ['search_docs', 'read_page']
        include_tools=['read_page'] -> ['read_page']

MCPTools now requests pages until the server returns no next_cursor, and it checks include_tools and exclude_tools only after it has every page. The cursor is opaque, so an empty or repeated cursor still leads to another request. Discovery stops at 250 pages, the same default budget as the FastMCP Client, and raises a RuntimeError at that limit so a partial catalog never registers. If a later page fails or discovery gets cancelled, the toolkit keeps the functions it already had.

Connections that MCPTools opens itself go through the FastMCP Client, which returns the complete list in one call, so that path works as before.

See the cookbook, and learn more about MCPTools configuration in the documentation.