One thing I’ve been thinking about:
Should Code Mode live purely in the MCP client, or could there be real value in having MCP servers also expose a “code mode” execution surface themselves?
For some mcp servers, having domain-specific batch operations or embedded execution could unlock more efficient workflows.
We have a platform (Gram) which is serverless MCP. Now users can launch MCP servers that only have 3 tools, search, describe, execute. Gram handles the tool routing, and execution (acting sort of like the sandbox). So the client only ever needs to know about those three tools. Everything else is abstracted.
One thing I’ve been thinking about: Should Code Mode live purely in the MCP client, or could there be real value in having MCP servers also expose a “code mode” execution surface themselves?
For some mcp servers, having domain-specific batch operations or embedded execution could unlock more efficient workflows.
We literally just did something like this on Gram, not exactly what you're saying, but it's pretty close... https://www.speakeasy.com/blog/how-we-reduced-token-usage-by...
We have a platform (Gram) which is serverless MCP. Now users can launch MCP servers that only have 3 tools, search, describe, execute. Gram handles the tool routing, and execution (acting sort of like the sandbox). So the client only ever needs to know about those three tools. Everything else is abstracted.
This is an interesting approach and I just saw your post on reddit too, good stuff!