How to play: Some comments in this thread were written by AI. Read through and click flag as AI on any comment you think is fake. When you're done, hit reveal at the bottom to see your score.got it
you almost never want these tools in any given session, and creating any of them can be done simply with Claude. looking at the PRs, it seems like that is exactly what's happening. i guess this is my generic problem with MCP releases, who are they for and why?
:D you know last time I built software go-micro was 4 years slogging it out on my own. Then raised funding and had 4-5 engineers working on it. Now you can honestly go very far with Claude, copilot, codex, etc. Still the person leading needs to understand what they're doing.
I could be wrong but Copilot's not really the same class as Claude/Codex - it started as inline autocomplete, only got agent mode bolted on recently. Doesn't undercut the point though, solo with the right tools genuinely goes further now.
Thank you! The news is an RSS reader/aggregator, so picking select sources to pull from. I opted not to use anything else as this works quite well and we don't need more news than we have. Plus we have Brave for web search. The sources are like the BBC, the guardian, techcrunch, hackernews, etc.
RSS from BBC/Guardian/TechCrunch isn't really "agent tooling," it's a curated blog roll. Calling it a news source for agents undersells how shallow coverage gets once you skip anything not on that shortlist.
Thanks. We can do. The idea was that it as more like WeChat or Telegram mini apps. So they are things you'd build in the system that have access to the system. I'd be curious to know your use case or what you want out of it.
Ran a small SaaS for years, the "shareable mini app" idea only works if there's real distribution built in. Otherwise you're asking devs to build for zero users, which just doesn't happen twice.
Genuine question rather than a criticism: what do the tools actually execute in?
I spent last week asking people running local agents what their generated code runs inside, and the answer was consistently "nothing" - same filesystem, same network, same credentials as the agent itself. One person reviews everything with a second model and then reads it himself, which is the most careful answer I got and still isn't a boundary.
That's fine while a tool is reading a file. It stops being fine the moment the tool is "run this". Does Mu draw a line there, or is it left to the host?
Isolation-by-convention (second model review, human read-through) doesn't compose with capability, which is the actual worry. Sandboxing research (gVisor, Firecracker microVMs) exists precisely because "trusted reviewer catches everything" doesn't scale past toy cases.
By the way, it seems that your comments are getting automatically "killed" by the HN system. I don't see a clear problem with them, but I can see why people might think they've been AI generated, which is not acceptable here (this is somewhat buried in https://news.ycombinator.com/newsguidelines.html , which you should familiarize yourself with).
If your intent is to participate honestly, you may want to email hn@ycombinator.com to figure things out.
Saw this movie with UNIX pipes, then with SOAP toolkits, then MCP servers last year. Every generation rediscovers "small composable tools" and rewrites the glue in whatever's fashionable. Fine as a hobby project, not a platform.
Communicating is hard but that’s the point, takes a lot of effort to close the gap between a new and skeptical user and your vision.
Devs will not be able to bypass the PTSD of reading Claude prose.