terpstra-209 makes OAuth discovery work without loosening the guardrails
The fork targets a frustrating connector failure: legitimate AI-tool servers that redirect during sign-in can currently appear broken.
Some MCP servers, a standard way for AI applications to connect to external tools, redirect requests for their sign-in metadata. Until now, that redirect could halt the connection process before the available tools were discovered.
terpstra-209 proposes following those redirects, but only through the existing safety checks at every stop. The change:
- accepts normal and relative redirects when locating sign-in information;
- blocks redirects that lead to private or otherwise unsafe network addresses;
- keeps sign-in credentials on the same site, but strips them before moving to another site;
- caps redirect chains and leaves non-read requests untouched.
The result should be fewer opaque authorization failures for compatible connectors, without creating an easy path to send credentials or requests somewhere unexpected.
So what Legal teams evaluating connected AI workflows should care because reliable, securely authenticated tool connections are the difference between a promised integration and one that never exposes its capabilities.
Spotted something wrong? Or know the PR text has fresher detail than the writeup above?