MCP and tool discovery

Tools you did not hard-code

Hard-coding a tool list works until you have four agents and eleven tools, at which point every agent needs redeploying to gain a capability. MCP — the Model Context Protocol — standardises the alternative: a server advertises the tools it offers, a client asks for them at startup, and the agent works with whatever it is given.

The protocol details are not the interesting part. The interesting part is what changes: the tool list becomes runtime data from an external source, and everything you assumed about it stops being guaranteed.

• Two servers can advertise a tool with the same name. One will shadow the other.
• A server can advertise a tool your agent must not be allowed to call.
• The list can change between runs, so a name that worked yesterday may be gone.

So a client that merges tool lists must resolve collisions deliberately rather than letting a later entry silently win.

Your turn: write merge_tools(servers) where servers is a list of (name, tools) pairs. Namespace each tool as server.tool so two servers can offer search without one erasing the other, and return the merged dictionary.

def merge_tools(servers):
    """Merge tools from several servers, namespaced as 'server.tool'."""
    return {}