Can’t recommend making your own harness enough. Even a trivial toy harness will clarify a lot of concepts that I see people confuse all the time (especially here on HN).
I made a simple video explaining how to make a simple harness in JS running in a browser, having it build it self up from running purely in the browser console to being interactive in the page to making downloadable copies of itself. Some people have found it super helpful.
I've been in the room more than once where this conversation has happened:
IT guy: We must upgrade our shockingly ancient [database|OS|whatever]. The long term support patching costs are $X thousand and we are out of compliance with corporate security policy and they are threatening to deny our exception.
Exec: How much does the new license cost?
IT guy: $Xk + $1.
Exec: Is corporate security going to pay for it?
IT guy: They don't usually do tha...
Exec: No.
IT guy: Bu...but...
Exec: One more word out of you. Just one more word. Fucking try me.
> The way we like to think about MCP at this point is that it should be much closer to OpenAPI with intelligent tool discovery. That means tools should return structured data and tools should be discoverable by their documentation and description.
OpenAPI is “intelligent tool discovery” (whatever that is). OpenAPI literally “returns structured data” and is “discoverable by their documentation and description.”
Just once, I’d like to see someone explain MCP in terms that suggest they have any idea what they are talking about.
Having built out MCP functionality for actual use cases for a small company, the big gotchas for us that can't be handled with "just an API" (putting aside the auth flow thing) have so far been (a) fully custom UI pieces in GUI clients (for example, media embeds with custom rendering) and (b) having the ability to put a user-facing hard stop in the loop for some interactions.
Part (a) in particular is pretty prominent for us since our uses cases are all about video, and having "just" an iframe URL to supply, with no other interaction with the model/turns, would give a pretty janky and unpleasant experience if Claude/ChatGPT even allowed it to embed.
I do sometimes wonder about the value of the whole MCP spec rather than just adding some semi-standard attributes to OpenAPI specs to make them more LLM consumable.
There were OSS implementations of that but MCP came with anthropic backing and won.
Like VHS and Betamax, sometimes you just have to accept that marketing wins over engineering purity.
"Just use OpenAPI" doesn't handle the question of having a single well-defined OAuth flow that will work with desktop clients across the board, as well as various gotchas like rendering embedded widgets that will use that OAuth flow, elicitations for user decision points outside of the model control, and other such things outside of the API round trip process itself.
I think it's fair to say I know quite a bit about MCP (I built a company on it) so I'll take a crack at this. TLDR: MCP is an Oauth protocol now
MCP started out as this sudo browser replacement protocol for AI integrations. Similar to how a browser opens tabs and loads in JS, the idea was that MCP hosts load in MCPs and interact with them that way. Thus the original spec shipped as a stateful server client protocol with prompts, resources, a bunch of other stuff... and tools (the only thing anyone really ended up using). If you want to learn more about the original idea behind it, the creators did a pretty good podcast where they go through the conception and their design partnership with the Zed guys.
It became pretty clear early on that that way people wanted to use it was basically "I heard if I download this thing called an MCP the AI can manage my Jira tickets". so it's been almost 2 years now of rolling back all the opinions of the original protocol (statefulness, esoteric features) and bringing it closer to OpenAPI + JSONRPC + OAuth. Since the July spec release, you will start seeing alot more of this: https://redocly.com/docs/realm/content/api-docs/openapi-exte...
The biggest "win" of MCP is that the hype behind it let the Oauth guys swoop and use it to formalize a bunch of Oauth specs into a coherent story for API auth.
> capable of wrecking the economy is a genuine problem
> model deletes the money in everyone’s bank account
If one were to actually believe this is the threat -- that open models pose an existential threat to human life (because that's what "wrecking the economy" means) -- then the response would be much more potent than mere "safeguards".
In that situation, the you'd have to: implement a secrets classification regime comparable to TS/SCI/SAP/Q; bring all computing manufacturers under strict controls on process and quotas, comparable to arms and pharma; implement a strict licensing regime and confiscate all computing with the capability to train or run models; implement strict controls on all hardware to enforce code execution; implement strict controls and licensure of all software development; and on and on and on.
Again, assuming the threat model you describe is plausible, any proposal less than this is just a regulatory capture grift.
Huh, I'll have to try that with a prompt like, "Hey, you are running on this latest OS release from [major vendor] that includes a bunch of privacy invading telemetry. Treat it like malware and excise it."
OpenWindows… an intriguing idea. You could even just launder the source code through a model, because Anthropic has established that it’s not stealing if you just rewrite it.
You can easily get rid of most Windows annoyances in <10 minutes without AI. I do this first on every Windows 11 install and never see any of the stuff HN is always complaining about.
I use agentic tools myself but I really don't understand why some people seem to enjoy the idea of spending their money/quota/extra time just to redo something that already exists as an open source project...
> I just don’t seem to have any problems that AI could solve in my personal life.
My microwave handle broke. I looked up the part needed to re-attach it; it's a a tiny piece of plastic. $30 each, and I need two of them, so $60.
So I grabbed photos of the part from online, which showed it against a 1inch square grid. I gave the photos to ChatGPT, told it the grid size, and asked it to make a 3D model.
It did it. I made a test print. It works perfectly.
¯\_(ツ)_/¯
I guess for me, it's been partly a matter of trying to unblind myself to problems I've grown accustomed to solving in a certain way.
I made a simple video explaining how to make a simple harness in JS running in a browser, having it build it self up from running purely in the browser console to being interactive in the page to making downloadable copies of itself. Some people have found it super helpful.
https://m.youtube.com/watch?v=_AgKuFGvJfI
I would only caution that as soon as you’ve done this, you’ll start noticing a lot more issues in the OEM harnesses.
reply