Client Side Tool Execution

[andrewlayer]

Pre-submission Checklist

• I have verified this would not be more appropriate as a feature request in a specific repository
• I have searched existing discussions to avoid duplicates

Your Idea

One of my favorite MCP features—one that no one else currently supports—is sampling. When I was building VSCode extensions using their Chat API, they had a similar concept: extensions could initiate inference using the client user’s own LLM API token. MCP’s sampling feature works in a comparable way.

Lately, I’ve been thinking about a related feature, but instead of LLM inference, it is tool execution.

The concept is this: during the MCP server’s initiation phase, the client (the one opening the connection) declares the tools it has available. The server could then issue ClientToolExecution notifications, effectively asking the client to execute tools on the server's behalf. This is particularly interesting in cases like Cursor, where the client already includes powerful built-in tools—like CLI access or file search. Rather than replicating those in a local MCP server, it would be ideal to simply leverage what the client already provides.

Pros:

• Avoids duplicating functionality unnecessarily
• Client-side tools are often more specialized and performant than server-side equivalents
• Servers can be highly configured and customized (They can basically be their own agents)

Cons:

• Unclear where execution responsibility lies—should the client or server run the tools?
• How should tools from other MCP servers be handled? For example, should something like the filesystem server expose its tools to other servers?
• Who is ultimately responsible for orchestrating tool usage?

Scope

• Protocol Specification
• SDK Features
• Documentation
• Developer Experience
• Other