Hacker News Reader

Comment by jadar | Hacker News Reader

I don’t want to undermine the author’s enthusiasm for the universality of the MCP. But part of me can’t help wondering: isn’t this the idea of APIs in general? Replace MCP with REST and does that really change anything in the article? Or even an Operating System API? POSIX, anyone? Programs? Unix pipes? Yes, MCP is far simpler/universal than any of those things ended up being — but maybe the solution is to build simpler software on good fundamental abstractions rather than rebuilding the abstractions every time we want to do something new.

View on HN

MCP is not REST. In your comparison, its more that MCP is a protocol for discovering REST endpoints at runtime and letting users configure what REST endpoints should be used at runtime.

Say i'm building a app and I want my users to be able to play spotify songs. Yea, i'll hit the spotify api. But now, say i've launched my app, and I want my users to be able to play a song from sonofm when they hit play. Alright, now I gotta open up the code and do some if statements hard code the sonofm api and ship a new version, show some update messages.

MCP is literally just a way to make this extensible so instead of hardcoding this in, it can be configured at runtime

so... is this OpenAPI then?

Basically, yes. But with much more enthusiasm!

OpenAPI doesn't have a baked in discoverability mechanism. It isn't compatible with LLMs out of the box. It is a lower level abstraction. I don't want to write a blob of code that talks to an Open API service every time I want to do something with an LLM.

>OpenAPI doesn't have a baked in discoverability mechanism.

Well, Swagger was there from the start, and there's nothing stopping an LLM from connecting to an openapi.json/swagger.yaml endpoint, perhaps meditated by a small xslt-like filter that would make it more concise.

HATEOAS was supposed to be that.

https://en.wikipedia.org/wiki/HATEOAS

heh, there was a good convo about HATEOAS and MCP on HN awhile back:

* https://news.ycombinator.com/item?id=43307225

* https://www.ondr.sh/blog/ai-web

Wait was it? HATEOAS is all about hypermedia, which means there must be a human in the loop being presented the rendered hypermedia. MCP seems like it's meant to be for machine<->machine communication, not human<->machine

I agree that HATEOAS never made sense without a human in the loop, although I also have never seen it be described as such. IMO that’s an important reason why it never gained useful traction.

There is a confused history where Roy Fielding described REST, then people applied some of that to JSON HTTP APIs, designating those as REST APIs, then Roy Fielding said “no you have to do HATEOAS to achieve what I meant by REST”, then some people tried to make their REST APIs conform to HATEOAS, all the while that change was of no use to REST clients.

But now with AI it actually can make sense, because the AI is able to dynamically interpret the hypermedia content similar to a human.

My understanding was that the discoverable part of HATEAOS was meant for machine to machine. Actually all of REST is machine to machine except in very trivial situations.

Not sure I'm understanding your point in hypermedia means there is human in the loop. Can you expand?

H in HATEOAS stands for "hypermedia". Hypermedia is a type of document that includes hypermedia controls, which are presented by the hypermedia client to a user for interaction. It's the user who makes decision what controls to interact with. For example, when I'm writing this comment, HN server gave a hypermedia document, which contains your comment, a textarea input and a button to submit my reply, and me, the human in the loop, decides what to put in it the input and when to press the button. A machine can't do that on its own (but LLMs potentially can), so a user is required. That also means that JSON APIs meant for purely machine to machine interactions, commonly referred to as REST, can't be considered HATEOAS (and REST) due to absence of hypermedia controls.

Further reading:

- https://htmx.org/essays/how-did-rest-come-to-mean-the-opposi...

- https://htmx.org/essays/hateoas/

So that's not my understanding. Hypermedia, as I understand it, are embedded links in responses that present possible forward actions.

They are structured in a way that machine program could parse and use.

I don't believe it requires human-in-the-loop, although that is of course possible.

HTML is a hypermedia format, the most widely used, and it's designed mainly for human consumption. Machines parsing and using something is too broad an idea to engage with meaningfully: browsers parse HTML and do something with it: they present it to humans to select actions (i.e. hypermedia controls) to perform.

Your understanding is incorrect, the links above will explain it. HATEOAS (and REST, which is a superset of HATEOAS) requires a consumer to have agency to make any sense (see https://intercoolerjs.org/2016/05/08/hatoeas-is-for-humans.h...)

MCP could profitably explore adding hypermedia controls to the system, would be interesting to see if agentic MCP APIs are able to self-organize:

https://x.com/htmx_org/status/1938250320817361063

I've programmed machines to use those links so I'm pretty certain machines can use it. I've never heard of the HTML variation but so will have a look at those links.

how did the machines you programmed react to new and novel links/actions in the response?

New are fine, Novel need catered for. I have left a fuller explanation to a sibling comment.

Right, so that means that the primary novel aspect of REST (according to the coiner of that term, Roy Fielding), the uniform interface, is largely wasted. Effectively you get a level of indirection on top of hard-coded API response endpoints. Maybe better, but I don't think by much and a lot of work for the payoff.

To take advantage of the uniform interface you need to have a consumer with agency who can respond to new and novel interactions as presented in the form of hypermedia controls. The links above will explain more in depth, the section on REST in our book is a good overview:

https://hypermedia.systems/components-of-a-hypermedia-system...

> I've programmed machines to use those links so I'm pretty certain machines can use it

I'm curious to learn how it worked.

The way I see it, the key word here is "programmed". Sure, you read the links from responses and eliminated the need to hardcode API routes in the system, but what would happen if a new link is created or old link is unexpectedly removed? Unless an app somehow presents to the user all available actions generated from those links, it would have to be modified every time to take advantage of newly added links. It would also need a rigorous existence checking for every used link, otherwise the system would break if a link is suddenly removed. You could argue that it would not happen, but now it's just regular old API coupling with backward compatibility concerns.

Building on my previous example of hn comments, if hn decides to add another action, for example "preview", the browser would present it to the user just fine, and the user would be able to immediately use it. They could also remove the "reply" button, and again, nothing would break. That would render the form somewhat useless of course, but that's the product question at this point

Yes, most of your observations about limitations are true. That doesn't mean it's not a useful technique.

This is a reasonable summary of how I understand it: https://martinfowler.com/articles/richardsonMaturityModel.ht...

This aligns with how I've seen it used. It helps identify forward actions, many of which will be standard to the protocol or to the domain. These can be programmed for and traversed, called or aggregated data presented using general or specific logic.

So New actions can be catered for but Novel actions cannot, custom logic would require to be added. Which then becomes part of the domain and the machine can now potentially handle it.

Hope that helps illustrate how it can be used programmatically.

If you have machine <-> machine interaction, why would you use HTML with forms and buttons and text inputs etc? Wouldn't JSON or something else (even XML) make more sense?

Hypermedia isn't just for human consumption. Back in the 90s, the Web was going to be crawled by "User Agents": software performing tasks on behalf of people (say, finding good deals on certain items; or whatever). Web browsers (human-driven interfaces) were the most common User Agent, but ended up being a lowest-common-denominator; the only other User Agents to get any widespread support were Google crawlers.

Feels like segment.com but for calling APIs rather than adding libraries to the frontend.

Now make the segment for MCPs ;p

MCP is a JSON RPC implementation of OpenAPI, or, get this, XML and WSDL/SOAP.

WSDL triggered me ha. I'm afraid you're right

>Alright, now I gotta open up the code and do some if statements hard code the sonofm api and ship a new version, show some update messages.

You will need to do that anyway. Easier discovery of the API doesn't say much.

The user might want a complicated functionality, which combines several API calls, and more code for filtering/sorting/searching of that information locally. If you let the LLM to write the code by itself, it might take 20 minutes and millions of wasted tokens of the LLM going back and forth in the code to implement the functionality. No user is going to find that acceptable.

Can't you just build a simple REST that takes this abstraction of plugging in different song providers away?

That only works if you let the LLM do the interpretation of the MCP descriptions, in the case of TFA the idea was to use MCP without LLM, which is essentially same as any old API.

You can use MCP to dynamically call different services, without ever having to use an LLM to decide.

With an LLM it would go

List MCP Tools -> Get User prompt -> Feed both into LLM -> LLM tells you what tools to call

You could skip the LLM aspect completely and get all tools and let the user at runtime pick the tool that "playsSong" for example

The main difference between MCP and Rest is that MCP is self described from the very start. REST may have OpenAPI, but it is a later addon, and we haven't quite standardised on using it. The first step of exposing an MCP is describing it, for Rest is is an optional step that's often omitted.

But you're describing it in a way that is useless to anything but an LLM. It would have been much better if the description language had been more formalized.

The description includes an input and output json schema.

Only input, not output.

https://modelcontextprotocol.io/specification/2025-03-26/ser...

You're not looking at the latest version. They added output schemas.

Thank you!

> It would have been much better if the description language had been more formalized.

To speculate about this, perhaps the informality is the point. A full formal specification of something is somewhere between daunting and Sisyphean, and we're more likely to see supposedly formal documentation that nonetheless is incomplete or contains gaps to be filled with background knowledge or common sense.

A mandatory but informal specification in plain language might be just the trick, particularly since vibe-APIing encourages rapid iteration and experimentation.

isn't also SOAP self described?

And gRPC with reflection, yeah?

and GQL with reflection?

JSON-LD?

When I read about MCP the first time and saw that it requires a "tools/list" API reminded me of COM/DCOM/ActiveX from Microsoft, it had things like QueryInterface and IDispatch. And I'm sure that wasn't the first time someone came up with dynamic runtime discovery of APIs a server offers.

Interestingly, ActiveX was quite the security nightmare for very similar reasons actually, and we had to deal with infamous "DLL Hell". So, history repeats itself.

Is it "self-described" in the sense I can get a list of endpoints or methods, with a human- (or LLM-) readable description for each - or does it supply actual schemata that I could also use with non-AI clients?

(Even if only the former, it would of course be a huge step forward, as I could have the LLM generate schemata. Also, at least, everyone is standardizing on a base protocol now, and a way to pass command names, arguments, results, etc. That's already a huge step forward in contrast to arbitrary Rest+JSON or even HTTP APIs)

For each tool you get the human description as well as a JSON schema for the parameters needed to call the function.

You're getting an arbirary string back though...

how else would you describe an arbitrary tool?

In my mind the only thing novel about MCP is requiring the schema is provided as part of the protocol. Like, sure it's convenient that the shape of the requests/response wrappers are all the same, that certainly helps with management using libraries that can wrap dynamic types in static types, but everyone was already doing that with APIs already we just didn't agree on what that envelope's shape should be. BUT, with the requirement that schema be provided with the protocol, and the carrot of AI models seamlessly consuming it, that was enough of an impetus.

> the only thing novel about MCP is requiring the schema is provided as part of the protocol

You mean, like OpenAPI, gRPC, SOAP, and CORBA?

You can’t connect to a gRPC endpoint and ask to download the client protobuf, but yes.

It's not enabled by default, but you can --- gRPC Reflection:

* https://github.com/grpc/grpc-java/blob/master/documentation/...

* https://grpc.io/docs/guides/reflection/

You can then use generic tools like grpc_cli or grpcurl to list available services and methods, and call them.

Where is the mandatory human readable prose description of the purpose of the tool in any of those specs. It isn't. Also the simplicity of JSON interface descriptions is key.

The main difference between MCP and REST is `list-tools`.

REST APIs have 5 or 6 ways of doing that, including "read it from our docs site", HATEOAS, OAS running on an endpoint as part of the API.

MCP has a single way of listing endpoints.

WSDL + XML API's have been around since 1998.

OpenAPI, OData, gRPC, GraphQL

I'm sure I'm missing a few...

Where is "list-tools" in any of those low level protocols?

All of them already provide an IDL with text descriptions and a way to query a server's current interface, what else do we need? Just force those two optional features to be required for LLM tool calls and done.

Is there anything stopping generic MCP servers for bridging those protocols as-is? If not, we might as well keep using them.

I don't know enough about OData, but:

- Introspection (__schema queries) for every graphQL server. You can even see what it exposes because most services expose a web playground for testing graohQL APIs, e.g. GitHub: https://docs.github.com/en/graphql/overview/explorer

- Server Reflection for gRPC, though here it's optional and I'm not aware of any hosted web clients, so you'll need a tool like gRPCCurl if you want to see how it looks in real services yourself.

- OpenAPI is not a protocol, but a standard for describing APIs. It is the list-tools for REST APIs.

In other words, there’s no commonly-used, agreed upon standard for creating APIs. The closest is REST-like APIs, which are really no more specific than “hit a URL and get some data back”.

So why are we all bitching about it? Programmatically communicating with an ML model is a new thing, and it makes sense it might need some new concepts. It’s basically just a wrapper with a couple of opinions. Who cares. It’s probably better to be more opinionated about what exactly you put into MCP, rather than just exposing your hundreds of existing endpoints.

I don’t think the comments here are complaining, they are pointing out that what’s being claimed as being new is not actually new

> The main difference between MCP and REST is `list-tools`.

> REST APIs have 5 or 6 ways of doing that

You think nobody's ever going to publish a slight different standard to Anthropic's MCP that is also primarily intended for LLMs?

Why would they? I'm sure the "Enterprise" folks are putting together some working group to develop ANSI-xyzzy standard for Enterprise operability which will never see the light of day.

Because they genuinely think it'll work better, because they think it will build brand awareness/moat, because they're upset MCP comes from a competitor

Damn, I just read this and it's comforting to see how similar it is to my own response.

To elaborate on this, I don't know much about MCP, but usually when people speak about it is in a buzzword-seeking kind of way, and the people that are interested in it make these kinds of conceptual snafus.

Second, and this applies not just to MCP, but even things like JSON, Rust, MongoDB. There's this phenomenon where people learn the complex stuff before learning the basics. It's not the first time I've cited this video on Homer studying marketing where he reads the books out of order https://www.youtube.com/watch?v=2BT7_owW2sU . It makes sense that this mistake is so common, the amount of literature and resources is like an inverted pyramid, there's so little classical foundations and A LOT of new stuff, most of which will not stand the test of time. Typically you have universities to lead the way and establish a classical corpus and path, but being such a young discipline, 70 years in and we are still not finding much stability, Universities have gone from teaching C, to teaching Java, to teaching Python (at least in intro to CS), maybe they will teach Rust next, but this buzzwording seems more in line with trying to predict the future, and there will be way more losers than winners in that realm. And the winners will have learned the classicals in addition to the new technology, learning the new stuff without the classics is a recipe for disaster.

Apis do not need to necessarily tell you everything about themselves. Anyone who has used poorly documented or fully undocumented apis knows exactly what I'm talking about here.

Obviously, for http apis you might often see something like an open API specification or graphql which both typically allow an api to describe itself. But this is not commonly a thing for non-http, which is something that mcp supports.

MCP might be the first standard for self-described apis across all protocols(I might be misusing protocols here but not sure what the word technically should be. I think the MCP spec calls it transport but I might be wrong there), making it slightly more universal.

I think the author is wrong to discount the importance of an llm as an interface here though. I do think the majority of mcp clients will be llms. An API might get you 90% of the way there but if the llm gets you 99.9% by handling that last bit of plumbing it's going to go mainstream.

honestly, yes - but MCP includes a really simple 'reflection' endpoint to list the capabilities of an API, with human readable docs on methods and types. That is something that gRPC and OpenAPI and friends have supported as an optional extension for ages, but it has largely been a toy. MCP makes it central and maybe that makes all the difference.

At a previous job most of our services supported gRPC reflection, and exploring and tinkering with these APIs using the grpc_cli tool was some of the most fun I had while working there. Building and using gRPC services in golang left a strong positive impression on me.

I had the same experience working with GQL :)

This is exactly what the author is saying. It is "the idea of APIs in general" that has suddenly become a fad under the guise of MCP, riding the AI wave. And it may well be a very imperfect way to build APIs, but if it eventually becomes the standard to the point where every app has to offer it, it's still "good enough" and would massively improve interoperability all around as a side effect.

One major difference is that MCP has discovery built into the protocol. There’s nothing in REST that informs clients what the API can do, what resources are available, etc.

My first thought as well. But maybe at least people wanting to plug their apps to their AI forces developers to actually implement the interface, unlike APIs that are mostly unheard of in general population and thus not offered?

Reply on news.ycombinator.com