Agno 3.0.2 turns each agent into a named MCP tool

Agno 3.0.2 changes how its agents talk to MCP clients. Before this release, an outside caller went through one dispatcher, something like run_agent(agent_id="chief"), with the agent ID passed as a string argument. Now that string becomes a tool name a client can see and call directly, which changes what shows up in a tool list and what an approval step has to reason about.

MCPConfig.tools now accepts Agent, Team, and Workflow instances directly. Each one shows up in the MCP client as its own named tool, so chief becomes a callable tool rather than a parameter value. Developers can also call component.as_tool(name=..., description=...) to control the name and description a client sees.

Toolkits get the same treatment at the method level. Each registered method on a Toolkit publishes as its own MCP tool, and the toolkit's existing enable_*, include_tools, and exclude_tools settings still control which methods get exposed. Return values now render as proper MCP content blocks, including a resource_link type for results that are just a URL rather than inline data.

The release also adds MCP title and behavior annotations. Anything exposed this way defaults to readOnlyHint=False, destructiveHint=True, and openWorldHint=True, per the Agno 3.0.2 release notes. An MCP client uses those hints to decide whether to ask a human before calling a tool or let an autonomous loop call it unattended. Defaulting everything to destructive is the cautious choice, but it puts the retuning work on whoever wired up the client, one flag at a time, per tool.