Why Can't I Access My MCP Server Due to Missing Authorization Header?

This topic was automatically generated from Slack. You can find the original thread here.

Hi, I can’t seem to access my MCP server. Previously, the URL had a unique uuid string, but now I have this;

I get an error saying it is missing authorization header

Where are you trying to connect to it?

Via Claude code MCP

Pipedream MCP now supports OAuth, so it’s far more secure than the previous implementation, and Claude Code supports OAuth

Have you tried that? Does that work?

i t works! thank you

nice feature

but wasn’t clear from the UI for the portal given my prior experience

Tried it just now tih the google sheets…for Claude Code, it doesn’t appear to work, It says connected, but then hits an error when I try to take an action, and not sure how tod ebug it

What’s the error you’re seeing?

i was getting authorization errors, starting over but will get the exact error

I think one thing was, with the same endpoint, should this be treated as a single MCP and I specify to the LLM which of the connected apps, or am I creating a distinct MCP server and naming the respective app but using the send endpoint?

The idea would be the former — you use Pipedream MCP to connect to >1 app

Ah… so the name of the app is sufficient for that MCP to reference its tools?

The LLM will also use the tool descriptions for tool identification / selection

What is the impact on context windows and performance? I have multiple tools that I connect via Pipedream and want to use with a single MCP server (but don’t have to; originally I was making a 1:1):

I get this message: Tool calling accuracy may decline with this many apps selected. Consider selecting fewer apps for better performance.

It doesn’t matter if 1 server has 100 tools or 100 servers each have 1 tool — at some point, too many tools will negatively impact the LLM’s context and ability to select the right tool. DOes that answer your question?

Right but if I specify different MCP servers even if they use the same end-point will that reduce the number of tools loaded for a given MCP server in claude?

Meaning: in a given call I don’t use all the tools so it looks like I should, in fact, separate into separate MCP servers even with the same URL (will that do it; or the fact I have the same pipedream end-point load them all every time)?