What about function name collisions?

[scottpersinger]

Pre-submission Checklist

• I have verified that this discussion would not be more appropriate as an issue in a specific repository
• I have searched existing discussions to avoid duplicates

Discussion Topic

Is there any guidance for how to avoid tool function name collisions? I note that the current GDrive tool exports a function called "search":

https://github.com/modelcontextprotocol/servers/blob/0e7d79c580da693091e7e9711a69132e85eb47c1/src/gdrive/index.ts#L133

which is HIGHLY likely to cause a collision down the road, haha. A related issue is that generic function names are likely to confuse your agent when they are mixed together. All the "search_file", "get_data", etc... functions are likely to only work reliably if your agent has that single tool attached.

My understanding is that function calling protocol at the moment doesn't support namespaces, so in the interim perhaps just providing guidance like:

Function names are recommended to include a prefix, like, "gdrive_search", rather than using more generic names.

would be a way to help ensure that tools can be used together by agents.

[jspahrsummers]

Yep, completely agreed. There's a somewhat related discussion here: #94

We may want to consider giving guidance to clients that they automatically use the server name for namespacing on their end.

[slimslenderslacks]

@jspahrsummers when building an agent loop pre mcp, it was sometimes useful to do tool selection before sampling. I guess you could kind of think of it as un-discovering the tools that were not relevant in the current thread. Do you anticipate an analogous requirement that impacts the mcp spec?
As an aside, it seems like 3.5 Sonnet is really quite good at dealing with pretty large numbers of available tools!

[jspahrsummers]

I don't know if it needs to be spec-ified, per se. I think we can give clients a lot of latitude in how they filter the tool list, like how you've done it.