Inspector client vs Cusror client

[jycouet]

I started building a small MCP server to experiment, and pretty quickly I got it working with the Inspector. It works really well — I'm able to list my tools and run them in the Inspector.

However, when I try to use it in Cursor (depending on the number of tools), I often get this error: The model returned an error. Try disabling MCP servers, or switch models.

Do you have any idea where I should look to debug this?

[letmaik]

Coverage of transitive supply chain provenance is not great at the moment. I've created https://letmaik.github.io/mcp-provenance-monitor/ to illustrate the current ecosystem state for a subset of MCP servers.

PyPI is in a slightly better state as their official GitHub Action automatically produces provenance when publishing packages from public CI like GitHub Actions. For npm, it is still opt-in via a flag.

I think it's too early to tell what packaging for MCP servers will eventually look like. The current conventions of running an npx/uvx or similar command won't last long I think. At the very least, client configuration schemas (https://github.com/modelcontextprotocol/modelcontextprotocol/issues/292) will have to change to be more declarative, for example "package": "pkg:pypi/[email protected]" (see purl). Like that, the client would be in charge of how to install and run the server, as well as any extra validation steps like checking provenance or applying company policies.