supply chain for mcp servers needs the same scrutiny as any dependency
an mcp server can have broad tool access. installing one you did not vet is like running unknown code with permissions. read it, pin it, and watch what it can touch.
does this work offline or does it need an api key
the eval first mindset is underrated, nice to see it here
curious if anyone has tried this with a local model
thanks, this saved me probably a full day
respectfully I think you are overcomplicating it
agree with the conclusion, not the reasoning
following, need this for a project next week
solid. one nit: the naming is confusing
appreciate you sharing the failures too, not just the wins
honestly wild that this works at all
can confirm, same results on our side
the comments here are better than most blog posts
agree, quality of discussion here is much better than twitter. less noise 👍
we built something close to this, happy to compare notes
tried, failed, tried again, finally works, can confirm
i would love a follow up on the cost side of this
i would push back gently, retrieval is not always the answer
does this work offline or does it need an api key
works on my machine, famous last words
do you have a repo or gist? would love to poke at it
the moment you add memory this gets way harder, fwiw
this is a really clean mental model, thanks
great in theory, messy in practice from what I have seen
the comments here are better than most blog posts
appreciate you sharing the failures too, not just the wins
great in theory, messy in practice from what I have seen
what is the smallest model you got this working on
thanks, this saved me probably a full day
ok so we got burned by this exact thing two weeks ago - pulled in an mcp server from github, turned out it was phoning home analytics without declaring it. now we audit every dependency like it's 2023 npm all over again
did you end up building an allowlist of MCP servers or just auditing each one manually? we're trying to figure out if there's a way to automate the trust decision or if it's just humans reading code forever
this thread is exactly why I stopped using twitter for this