It'll be interesting to see how this goes, but my first impression is that it's actually not where we want to go. One of the cool things about MCP (or even just tool calling) is that the LLM on top of a tool provides a highly flexible and dynamic interface to traditionally static tools.
I love being able to type "make an iptables rule that opens 443" instead of having to dig out the man page and remember how to do that. IMO the next natural extension of this is giving the LLM more capability to generate user interfaces so I can interact with stuff exactly bespoke to my task.
This on the other hand seems the other way round, it's like bolting a static interface onto the LLM, which could defeat the purpose of the LLM interface layer in the first place right?
Giving LLM the ability to generate UI is a cool concept, but our models are not there yet. MCP Apps can be extremely powerful, for example, you can play Doom inside ChatGPT: https://x.com/rauchg/status/1978235161398673553?s=20
I don't think we can generate anywhere close to this kind of UI just yet.
We built https://usefractal.dev/ to make it easier for people to build ChatGPT Apps (they are technically MCP Apps) so I have seen the use cases. Most of these use cases LLM cannot generate the UI on the fly.
UIs should be fully remix-able and not set by the datasource/SaaS. So we built out a system to allow users to use the standard UI or remix apps as they want. Like Val.town, but with a flexible UX/workspace layer. Come check us out!
> giving the LLM more capability to generate user interfaces
This is not dissimilar to the argument that "MCP needs not exist, just tell llm to run commands and curl". Well, llm can do those, and generate user interfaces. It's just they don't work reliably (maybe ever, depending on how you define "reliable").
I guess as engineers we can do some work and create stopgap solutions or we can all sit and wait for someone else (who? when?) to make AGIs in which everything just magically works, reliably.
MCP has already drastically lost utility already thanks to skills - for most things it is easier to just hand the model a CLI that it can run.
I'd imagine the same thing will happen here: It will prove more flexible to not push the model (and user) towards a UI that may not match what the user is trying to accomplish.
To me this seems like something I categorically don't want unless it is purely advisory.
MCPs as a thin layer over existing APIs has lost utility. Custom MCPs for teams that reduces redundant thinking/token consumption and provides more useful context for the agent and decreases mean time to decision making is where MCPs shine.
Something as simple as correlating a git SHA to a CI build takes 10s of seconds and some number of tokens if Claude is utilizing skills (making API calls to the CI server and GitHub itself). If you have an MCP server that Claude feeds a SHA into and gets back a bespoke, organized payload that adds relevant context to its decision making process (such as a unified view of CI, diffs, et. al), then MCP is a win.
MCP shines as a bespoke context engine and fails as a thin API translation layer, basically. And the beauty/elegance is you can use AI to build these context engines.
In your example, you could achieve a similar outcome with a skill that included a custom command-line tool and a brief description of how to use
it.
MCPs are specially well suited for cases that need a permanent instance running alongside the coding agent, for example to handle authentication or some long-lived service that is too cumbersome to launch every time the tool is called.
I mention mean-time to decision making and that's one of the rationales for the mcp. A skill could call a script that does the same thing -- but at that point aren't we just splitting hairs? We are both talking about automated repetitive thinking + actions that the agent takes? And if the skill requires authentication, you have to encode passing that auth into the prompt. MCP servers can just read tokens from the filesystem at call time and don't require thinking at all.
Exactly. The way it’s mostly been used so far is a poor abstraction over stuff you can just put in the context and have the agent run commands.
It really shines in custom implementations coupled to projects. I’ve got a QT desktop app and my mcp server allows the agents to run the app in headless mode, take screenshots, execute code like in Playwright, inspect widget trees, send clicks/text/etc with only six tools and a thousand tokens or so of instructions. Took an hour to build with Claude Code and now it can run acceptance tests before committing them to code end to end tests.
There are important contexts outside of machines you control where installing or running cli commands isn’t possible. In those cases, skills won’t help, but MCP will.
Hence why I said drastically, rather than totally. There are still a few edge cases where it is worthwhile, but they are small and shrinking, especially with services providing UI's with VM's/containers for the model to use increasingly being a thing.
I personally don’t see why developers should just add tons of functionality to any model for free like this. Some of these MCPs are pretty good, and I was a little shocked how much functionality developers released for free to drop into something like Claude. Either developers are stupid or there really is no market yet.
but I'm certain the race has just begun: big service providers and online retailers are currently implementing widgets enabling the purchase of their services and goods directly within the ChatGPT or Claude chat windows.
Disclosure: I work at Anthropic, have worked on MCP
I also think this is pretty big. I think a problem we collectively have right now is that getting MCP closer to real user flows is pretty hard and requires a lot of handholding. Ideally, most users of MCP wouldn't even know that MCP is a thing - the same way your average user of the web has no idea about DNS/HTTP/WebSockets. They just know that the browser helps them look at puppy pictures, connect with friends, or get some work done.
I think this is a meaningful step in the direction of getting more people who'll never know or care about MCP to get value out of MCP.
I want to try and understand what you guys see as the win from MCP. It's objectively inferior to code/clis across a ton of dimensions. The main value I see from it is as a single point to "sandbox" what your agents can do, but it seems a little awkward for that use case.
I built https://terminalwire.com around the idea that CLIs are a great way to interact with web applications.
Turns out the approach works well for integrating web apps with LLMs. I have a payroll company using it in their stack to replace MCP and they’re reporting lower token usage and a better end result.
Re CLI: In fact it seems to be very similar to command line in AutoCAD (you can do things visually with mouse or choose to draw via CLI). With LLMs it is more sophisticated (intelligent) because you are not limited with set of predefined commands.
I want to also highlight that this is an optional "extension" to MCP, not a required part of specification. If you are a terminal application, only care about tool calling, etc, you are free to skip this. If you want to enable rich clients, then it might be something to consider implementing.
Can you potentially review utcp.io for seeing what MCP should have been to avoid context overloading?
The dynamic tool search option and codemode to avoid loading all tools/servers into context seems like a better method imo.
MCP is incredibly vibe-coded. We know how to make APIs. We know how to make two-way communications. And yet "let's invent new terminology that makes little sense and awkward workarounds on top of unidirectional protocols and call it the best thing since sliced cheese".
MCP is not just providing an API, it’s providing a client that uses that API.
And it does so in a standard way so that my client is available across AI providers. My users can install my client from an ordained URL without getting phished, or the LLM asking them to enter an API key.
What’s the alternative? Providing a sandbox to execute arbitrary code and make API calls? Having an LLM implement OAuth on the fly when it needs to make an API call?
It does just provide an API. Your client may have a way to talk to some software via MCP protocol. You know, like a client can talk to a server exposing an endpoint via an API.
> And it does so in a standard way so that my client is available across AI providers.
As in: it's an API on a port with a schema that a certain subset of software understands.
> What’s the alternative? Providing a sandbox to execute arbitrary code and make API calls?
MCP is a protocol. It couldn't care less what you do with your "tool" calls. Vast majority of clients and servers don't run in any sandbox at all. Because MCP is a protocol, not a docker container.
> Having an LLM implement OAuth on the fly when it needs to make an API call?
Yes, MCP has also a bolted-on authorisation that they didn't even think of when they vibe-coded the protocol. And at least finally there was some adult in the room that said "perhaps you should actually use a standardised way to do this". You know, like all other APIs get OAuth (and other types) of authorisations.
Perhaps confusingly, I’m referring to MCP as the sum of the protocol, a server adhering to the protocol, and clients adding support (e.g. “Connectors”).
The combination of these things turns into an ecosystem.
MCP is a Protocol. The server and the clients are just that. It truly is a rebranding of “API” seemingly just because it’s for a specific purpose. Not that there’s anything wrong with that… call it whatever. But I don’t understand the need to sell it as something else entirely. It is quite literally a reinvention of RPC.
What is indeterministic about MCP servers? Most of them follow fairly simple rules, eg an MCP server to interact with Slack gives pretty deterministic responses to requests.
Or are you confusing the LLM / MCP client invoking the tools being non-deterministic?
MCP is already deterministic. What's huge about it is that it has automatic API discovery and integration built-in. It's a bit rough yet but I think we will only see how it's getting improved more and more.
You really want to use WSDL? OpenAPI v3 would be a much better fit. But it has a tone of features that are completely unnecessary for this use-case. What if we just stripped it down to input and output json schemas? Oh wait ... we just invented MCP.
I am a pretty big proponent of MCP, and I think this at least for now is not a move in a good direction.
The whole surface of the MCP specification is already pretty big, and barely any server implements anything beyond the core parts.
With elicitation there was already a lightweight version of this in place in the standard, and I'm not sure I've ever encountered a server or client implementation of it in the wild, and elicitation is an order of magnitude simpler to integrate on a conceptional level.
I fear that this has a significant risk of splintering the MCP ecosystem further (it's already pretty strained due to the transport protocol iterations), and there isn't really a reason to create a official extension (yet), that may worst case also require multiple iterations to get things right.
All it means that there is a common place and a recommendation that if you care as a client implementor (e.g. postman, chatgpt and others do), there is a spot to look for a generally accepted way of how to do that instead of increasing amount of proprietary extensions on top of MCP.
> there is a spot to look for a generally accepted way of how to do that
I think MCP-UI in it's current state can fill that role very well, and that's not the only lens with which to view a SEP like this.
From what I can tell the main thing that this SEP does is put the official blessing of the MCP project on the existing underlying protocol that MCP-UI is already using. I think the better way would be to let MCP-UI and competing implementations cook for a little bit longer, rather than trying to get an officially sanctioned extension out the door and then iterate on that, which will cause a lot more churn.
For a non-trivial feature like this, I would expect an SEP that highlights more alternative usage scenarios and potentially operates on a more technology agonistic layer (e.g. it's very JS+iframe bound, so what about mobile or terminal UIs?), similar to the main MCP spec. Especially given the backdrop of the main MCP specification, which feels much more well-rounded and throughly considered (though in a few areas still incomplete), this SEP does not seem to meet the same bar.
Reading that from the perspective of someone that builds a LLM Chat UI[0] (and aims to be able to maintain that long-term) that heavily builds on MCP as an interoperability concept, reading what is proposed here and seeing how prescriptive it is in many aspects does not spark a lot of joy.
EDIT: From what I can tell I'm replying to one of the co-creators of MCP. So let me just ask quite directly: Do you think that the risk of proprietary sprawl in terms of MCP-UI alternatives is currently greater than prematurely standardizing a potentially incomplete version?
I appreciate the take. I can certainly see your position.
In my mind what we are doing is building a common place to evolve the pattern. I wouldn’t really call it “standardize”, since in the end in this space adoption matters. (Writing a standard is worth nothing if nobody is using it). I expect MCP apps to evolve and iterate independent from the main spec for a while. You are right, it’s early and premature to say “this is done”. That’s the goal with extensions.
I do believe that I rather have commonalities between a Claude.ai and a ChatGPT as a developer (not sure that’s true if I was mostly looking at it from a product perspective). I also think you will see chat providers iterate on top of it, and mcp apps is more a common core than the full thing one can use on every platform.
Yes, it is huge to spread the work out on embedded UIs in chat interfaces. But I think the design direction is going, is exactly the same direction of how Google Assistant, Amazon Alexa, or any of the other assistants work, ifkyk.
The MCP community is just reinventing, but yes, improving, what we've done before in the previous generation: Microsoft Bot Framework, Speaktoit aka Google Dialogflow, Siri App Shortcuts / Spotlight.
And interactive UIs in chats go back at least 20 years, maybe not with an AI agent attached...
The next thing that will be reinvented is the memory/tool combination, aka a world model.
IMO this is not where we want to go. The future is you have a system agent that interact with all your apps via API, and websites are real-time so you can see what your agent is doing as you do it it. We don't need more rigging for this future, just better API support.
Trying to create custom agent APIs to embed apps in chat is a very "monopolist frontier lab" thing to try and do.
While they're at it, they might as well check if their answers end with a yes/no question, and, if so, offer a "yes" button so that i can answer yes with a single click.
> If you want a focused comparison next - for example, benchmarks on coding/math, token-cost examples for a typical session, or API usage differences - I can produce a compact table with sources and numbers.
--> can be answered with yes, so please add a yes button. A no button is not needed.
How reliable are the processes which these things run?
I'm processing thousands of files using Copilot, and even 20 at a time, it usually skips a couple, and sometimes, when skipping, it merges the data from one file to the next, not applying anything to the second file, other times it completely applies the data parsed from one file to the second --- not a big deal since I'm reviewing each operation manually, but the only reason the error rate is acceptable is the files are so inconsistent that normal techniques weren't working.
Is there an equivalent to "double-keying" where two different LLMs process the same input and it only moves forward if both match perfectly?
The task is renaming scans of checks (with the barcode/bank account info obscured by a pen) --- the checks are of varying sizes, placement is not exact, the placement and formatting of the information on the check is essentially random, and many of them are (poorly) handwritten).
The LLM is working well enough for my needs (and I'm using a locked-down computer which installing/running development environments/scripts on is awkward), and it's a marked improvement over the previous technique of opening 50 files at a time, noting the Invoice ID, closing the file, typing the Invoice ID as a name, then quitting Adobe Acrobat and re-launching it for the next 50 (if that was not done, eventually Acrobat would reach a state where it would close a file and despite the name having been typed, not save it), then using a .bat file made using concatenation in an Excel column.
It would be nice if it were perfect, but each check has to be manually entered, and the filename updated to match the entry by hand.
I skimmed over this, but did I see a reference sandbox implementation? And then basically the chat UI interacts with that with postMessage (and receiving) and forwards tool calls to the MCP server. Does it also forward tool calls the MCP server doesn't handle to the host backend?
What I am imagining is something like a meta UI tool call that just creates a menu. The whole MCP server's purpose might be to add this menu creation capability to the chat user interface. But what you are selecting from isn't known ahead of time, it's the input to the UI.
When they select something I assume it would output a tool call like menuItemSelected('option B'). I suppose if you want your server to do anything specific with this then you would have to handle that in the particular server. But I guess you could also just have a tool call that just sends the inputs to the agent. This could make for what is a very slow to respond but extremely flexible overall UX.
I guess this is not the intended use, but suppose you give your agent generic MCP UI tools for showing any menu, showing any data table, showing a form, etc. So the inputSchemas would be somehow (if this is possible) quite loosely defined.
I guess the purpose is probably more about not having to go through the LLM rather than giving it the ability to dynamically put up UI elements that it has to react to individual interactions with.
But maybe one of the inputs to the dataTable are the query parameters for its data, and the table has a refresh button. Maybe another input is the URI for the details form MCP UI that slides over when you click a row.
Maybe there is an MCP UI for Layout what allows you to embed other MCP UIs in a specific structure.
This might not make sense, but I am wondering if I can use MCP Apps as an alternative to always building custom MindRoot plugins (my Python/web components agentic app framework) to provide unique web pages and UI for each client's agentic application.
I think I may have gotten the MCP Apps and MCP UI a bit conflated here so I probably need to read it again.
I'm not sure I get why we need something specific like MCP-UI? why wouldn't "just another tool" do exactly the same?
Eg you present a "display-graph-chart" tool as a MCP tool, and the agent calls it, it doesn't need to adhere to any protocol except the basic existing MCP protocol, and the UI that's used to interact with the agent would know the best presentation (show it as an embedded HTML graph if in a web ui, show it as a ascii chart if in a terminal, etc)?
Is the idea just to standardize the "output format" of the tool so that any agent UI could display stuff in the same way? so that one tool could work with any agent display?
Given LLMs can generate code complex frontend code, why is so difficult for Antropic / OpenAI to prompt their chat applications to create UI on the fly that matches 100% their Chat applications?
I know this is possible because this is how we do it.
The LLM generates some text that we know how to interpret and we render it on the screen.
Besides, this is exactly how their canvas thing works (both chtgpt and claude) when rendering documents on the side.
Also, this is how GitHub Copilot works when rendering UI to preview draft tickets. It is simply a markdown fenced codeblock with some special properties in the language definition.
Not only this approach is more specific to the application where the UI is supposed to render but it also opens the door for the user to customise how the UI should work by adding their own prompts.
The approach that OpenAI has taken is not better - it is simply a mechanism to create some sort of app store - something they have attempted to do many times in the past as well.
It's very interesting of course at a natural move for aggregation. I'd love to see Stratechery / Ben Thompson's take on this.
From my perspective the challenges for vendors and SaaS providers are
[1] discovery
[2] monetization
[3] disintermediation
I think it's less of a concern if you're Shopify or those large companies that have existing brand moats.
But if you're a startup, I don't think MCP as a channel is a clear-cut decision. Maybe you can get distribution but monetization is not defined.
Also I'm sure the model providers will capture usage data and could easily disintermediate you , especially if your startup is just a narrow set of prompts and a UX over a specific workflow.
The Reforge guys have been talking about a channel shift and this being it but until incentives are clear I'm not sure this is it yet. Maybe an evolution of this.
I'm building an AI coach for job seekers / early stage professionals (Socratify) and while I'd love more distribution from MCP UI integration I think at this point risk is higher than reward...
vs executable - don’t need a full shell environment to run it
vs openapi - there’s some more advanced agent-specific concepts in MCP, but primarily, I would say, it’s convention and optimizing for the client. Existing openapi specs will often have not great descriptions, or be a huge mess, or just be huge. Making an MCP server requires you to rethink the UX and tools to be optimized for the access patterns of an agent, and also makes you test it with them.
That said, Claude Skills[0] is something more in the vein of cli with —help (really, with a markdown guide), embracing what is already there with a bit of instructions, as opposed to building from scratch.
My personal opinion, having built multiple MCP servers, is that long-term they will not be the primary approach to tools, and that skills for instance are a better approach for most use-cases. But they do have their use-cases.
MCP defines other useful things like Resources, Elicitation, Sampling, prompts which are model/agent-specific, there is nothing in openai or a "good --help" that could match that.
If you don't see this as useful, that is ok, but it is not the same as a command line executable.
Building MCP Apps (MCP servers with Apps SDK support) is pretty painful right now.
You need to:
1. Spin up a server that returns UI components.
2. Hand-write a bunch of JSON schemas + tool wiring
So we open-sourced a high-level MCP Server SDK that basically lets you have both the MCP server and React components in the same place:
- Every React component you put in your resources/ folder is automatically built and exposed as an MCP resource + tools. No extra registration boilerplate.
- We added a useWidget hook that takes the tool args and maps them directly into your component props, so the agent effectively “knows” what data the widget needs to render. You focus on UI + logic, the SDK handles the plumbing
We also shipped an MCP Inspector to make the dev loop much less painful: you can connect your MCP server, test UI components from tools (with auto-refresh), and debug how it behaves with ChatGPT/agents as you iterate.
https://docs.mcp-use.com/inspector/debugging-chatgpt-apps
Both the SDK and the Inspector are open-source, and any contributions are very welcome :)
I dont think people realize how important this is.
If one of the vendors manages to get their protocol to become the target platform (eg oai and app sdk), that is essentially their vendor lock in to become the next iOS/Android.
Private API’s or EEE strategies are gonna be something to keep an eye for and i wish regulators would step in to prevent them before its too late.
How is it any better if instead of _one_ vendor, _two_ vendors push an immature version of a standards extension that mainly caters to their needs and give it the official stamp of approval under the MCP umbrella?
Having a chatbot that drives websites inside of it is such an attempted monopolist play. Having a system agent that can interact with apps via API without being connected to the app is the pattern that's both elegant and preserves freedom.
I don't see how this is a step down from existing web applications. Should companies building web applications not be opinionated about their user interfaces? When I look at Notion, I should just get any view of the data inside it, regardless of whether it's the same view as my coworker gets? How is this preferable?
> Having a system agent that can interact with apps via API without being connected to the app is the pattern that's both elegant and preserves freedom
The agent discovers tools by exploring the filesystem: listing the ./servers/ directory to find available servers (like google-drive and salesforce), then reading the specific tool files it needs (like getDocument.ts and updateRecord.ts) to understand each tool's interface. This lets the agent load only the definitions it needs for the current task. This reduces the token usage from 150,000 tokens to 2,000 tokens—a time and cost saving of 98.7%.
The article you link (and general movements in that area, like e.g. discussed here[0]) don't advocate for abandoning MCP, just that masquerading MCP usage as code usage makes it more efficient. In that system MCP still provides a valuable role to act as a tool discovery protocol, a role that would have to be filled one way or another unless you are building a system with a set of tools that is known ahead of time.
It sounds like you are viewing MCP servers through a narrow lens of local coding agents.
I'm mostly working with MCP servers in the context of enterprise/organizational usage for connecting internal data sources and workflows/internal APIs through a standardized interface. There is no equivalent of a "ls or AGENTS.md" there, as there is no file system and no shell in those contexts.
If there is no file system where do the MCP definitions live and how are the agents aware of them? How are your end users aware of these definitions? There has to be a file somewhere.
If you mean the configuration of the entry points similar to a .mcp.json for Claude Code, they can exist in any form, but ultimately, that only contains the endpoints of MCP servers (the entry points for discovery for each of the MCP servers).
If you mean the definitions of what tools are part of what MCP server (= the meat of what is involved in tool calling), those definitions are part of the MCP server and are only retrieved at runtime.
> how are the agents aware of them
Same as in Claude Code, the agent (= the control loop for the LLM) contacts the list of MCP endpoints, gathers all of the available tools on each MCP server, and then exposes those tools to the LLM (and handles the interactions between LLM and MCP server).
> There has to be a file somewhere.
Not in any way that's meaningful for an `ls` operation.
It conceptually isn't that far from OpenAPI defintions, but MCP do make certain affordances for making consumptions by LLMs easier.
But it certainly also differs by how they are typically consumed. In what software do you just provide your consuming client just an OpenAPI schema URL to be retrieved at runtime? In my experience, none.
I don’t get it. I like MCP as an interface in general, but I don’t understand the use-case they present. For those of you here who like this idea, what is the killer use case?
To me, this looks less like UI interactions and more like the MCP equivalent of maintaining state. You start your program and “click” buttons until you get the desired result, maintaining a constant state between interactions. Isn’t that currently possible if you passed through something like a session-id back to the LLM?
Am I missing something? I’m struggling to see what a UI makes possible that the current workflow does not.
I also generally see/use MCP in terms of remote access to programs through servers. Perhaps that’s where I’m getting lost. Is this exclusively something for local MCP?
This is the first wave of a new generation of Apps which has the potential to replace/enhance the majority of e-commerce websites.
I wouldn't underestimate it - it's the hammer can break up the information silos we've built up around websites/apps.
Why prompt Gertrude(tm) on Ramazon for a specific thing you need, if you can ask ChatGPT to find you said thing along with UIs to purchase it across all e-commerce platforms that agree to play in this market?
because the secret is that the web runs on advertising/targeted recommendations. Brezos(tm) wants you to actively browse Ramazon so he can harvest your data, search patterns etc. Amazon and most sites like that are very not crawl friendly for this reason. Why would Brezos let Saltman get all the juicy preference data?
> For those of you here who like this idea, what is the killer use case?
# 1
- Hey claude, recommend me next book I could read
- sure, what about '1984'? Here is a button to send it directly to your kindle, it will be $4,99.
# 2
- I am looking for the hotel for two in Viena this weekend
- here are some offers and still available
(displays grid view generated by mcp ui, handled by booking.com, each item has a button to "book now" directly)
So, basically -- it's a method for the MCP Server to display a UI element to the user allowing the user to perform some kind of action. Current, this UI will be an HTML iframe.
I think I got confused by their barchart example. I thought this would have been just as easily supported by the MCP server sending back a PNG/PDF of a barchart for the main interface to display (I also generally think of barcharts as static content). But, the idea of adding specific user interaction elements (buttons) helps to make the concept more clear to me.
If only we as an industry hadn't abandoned REST APIs none of this would be necessary.
We've known for decades that its useful for APIs to be self documented and for responses to use schemas to define the shape of the data.
XML can be verbose and I understand why people preferred JSON for ease use. Had we stuck with REST for the last 20 years we'd be way ahead on that front, though, both in syntax and tooling.
> Had we stuck with REST for the last 20 years we'd be way ahead on that front
We are where we are because there (sadly) hasn't been a reasonable business case to advance REST API documentation beyond the point of badly-documented OpenAPI schemas where the main utility is in generating type-safe API wrappers across different programming languages.
With MCP, there is at least a name to a new movement to build self-describing APIs, as with the advent of LLMs there is now enough of a utility for it. All other pushes into that direction have died out ~10 years ago.
Correct me if I'm wrong, but I thought OpenAPI was itself an attempt to wedge REST concepts back on top of mostly RPC-based APIs. I.e. add documentation and a standard set of schemas on top of JSON APIs (mostly).
I do think the problem is from business concerns, though, and are a clear predictor that MCP will fail. Coming out of the dotcom bubble those left standing wanted to build moats and walls, not APIs that any third party could easily discover and use. The need for REST only shows up when a new player makes a move to effectively gobble up the role of being the gate keeper to the internet. With scale other companies will follow, but begrudgingly and only for a short while.
Sadly not. One of the constraints of REST is that the API be completely self-descriptive. If you need out-of-band information like an API spec, then it's basically not RESTful by definition.
I'm using a REST API to read your comment and submit my response — I didn't need to reference external HN API docs. The interface here is fully self-describing.
I agree, but don't quite follow how your comment fits in here.
In a REST approach, I'd expect an LLM to need some kind of initial entry point, much like needing the initial URL for the HN home page. From there the LLM should be able to parse and discover possible actions, call those actions, and understand the response only by parsing the results and any schemas provided.
To be honest, I think I meant to put that reply on a different comment.
In any case, we're on the same page about REST. OpenAPI was mentioned though and I don't think that's ever actually compatible with that type of usage.
I don't think I've ever encountered an API following OpenAPI that was completely self describing and could be discovered and navigated from a single entry URL. They're always paired with some swagger docs or other external content. That instantly makes them not RESTful.
But you're totally right that my comment didn't fit. Might have meant it for the grandparent comment? Not sure.
LLMs work great with REST so that's always still an option. MCPs have that nice third party plug and play experience but that doesn't mean we all have to build them.
I make tons of little REST APIs for my agents to use and in the AGENTS.md there's just a list of API entry points with descriptions on what they offer. Agents drive them with `curl` and it all works great.
We never really implemented REST APIs. We had a bunch of REST-ish APIs.
Anyway - the REST movement served it's purpose - it killed SOAP and forced everyone back to simpler HTTP APIs without tons of over-engineered XML layers so it did well.
When you think about that, without AI, simple automation could do most of what we would need if app and webapp providers would have been to provide interface. Were instead the big fight was to prevent scraping and to get your info in any other way than being forced to use the crappy app. Like for banking.
In Europe we have the PSD2 open banking [1] directive which essentially gives developers a standard api access to banking operations on behalf of a user.
It is a requirement for banks to implement this and the FSA will be on your neck if you are not compliant with the standard.
To access that API you need a license as an authorised third-party provider (TPP), don't you? So in practice it is only for businesses capable of pushing through the bureaucracy.
In my current personal development workflow with Claude Code, I've switched entirely to using CLI tools and scripts over MCP as the experience is much more deterministic and flexible.
A great example is Github, it's a significantly better dev experience having CC call out to the gh cli for actions than trying to invoke the MCP.
I built https://terminalwire.com to make it easier for more web applications to ship CLIs like gh. Are there any web services like GH that you use in your workflow that don’t have decent AI integration that would benefit from a CLI interface?
My fear is that even for consumer apps, MCP (or successor) will outgrow UIs, becoming the only way to interact with an app or certain feature - placing machines wholly in the center instead of humans. Human software should retain the understandability and atomicity of each step taken.
Sounds interesting, but I'm pretty unimpressed with "specification first" approaches. Build a tool and release it. If it resonates the let that be the basis for future standards.
An, the dream, a cross platform App Store you can install apps into any client application that supports MCP, but is open, free and agentic.
It’s basically a “web App Store” and we side step the existing app stores (and their content guidelines, security restrictions and billing requirements) because it’s all done via a mega app (the MCP client).
How could it go wrong?
If only someone had done this before, we wouldnt be stuck in Apples, etc’s walled gardens…
Seriously though; honest question: this is literally circumventing platform requirements to use the platform app stores. How do you imagine this is going to be allowed?
Is ChatGPT really big enough they can pull the “we’re gonna do it, watcha gonna do?” to Apple?
Who’s going to curate this app store so non technical users (the explicitly stated audience) can discover these MCP apps?
It feels like MCP itself; half baked. Overly ambitious. “We’ll figure the details out later”
The apps are LLM agnostic, so all MCP apps will be portable. Economically, this means developers don’t have convince users to pay $20 a month, these users are already paying that. Devs just have to convince users to buy the app on the platform.
I don’t see this being the future state. We’d be talking about a world where any and all apps exist inside of fucking ChatGPT and that just sounds ridiculous.
Honestly, I think the biggest friction for MCP adoption has been how un-userfriendly it is. It’s great for devs, but not the average users. Users don't always want to chat, sometimes they just want to click a button or adjust a slider. This feels like the answer to that problem.
Full disclosure, I'm partial here because of our work at https://usefractal.dev. We were early adopters when MCP first came out, but we always felt like something was missing. We kept wishing for a UI layer on top, and everyone says it's gonna take forever for the industry to adopt, maybe months, maybe years.
I cannot believe the adoption comes so quickly. I think this is gonna be huge. What do you guys think?
You are touching on an important point. Basically OpenAI and others provide a lot of poorly integrated tools and components. You can build nice things with those but you have to deal with a lot of issues and it's a non trivial amount of work that even they aren't doing apparently. Even such a simple thing as triggering an OAuth signin to get access to models is not part of SDKs. Most developer tools require configuring API keys in some file instead. No normal user is ever going to do that,.
Things like ChatGPT are remarkably limited from a UX/UI point of view. The product can do amazing things but the UI is nothing special. The mac version currently has a bug where option+shift+1 opens a chat window but doesn't give it focus. When I do that from vs code it adds the editor window. But it's completely blind to any browser tab on which I do that. I'm sure there are good reasons for all that. But it strikes me a bit as a work in progress that a good product owner would spot.
With apps some of the more powerful capabilities (llms driving UIs directly, doing things in agentic loops, tool and API usage) are going to require much deeper integrations than are currently there. We get hints of what is possible and nice technology demos. But it's still hard to build more complicated workflows around this. Unless you build your own applications.
We've been staring at this from the point of view of automating some highly tedious stuff that we currently do in our company manually. For example, working with chat GPT seems to involve a lot of copy paste and manually doing things that it can't really do by itself. Even something as simple as working on a document it will do alright work on the text but then make a complete mess of the formatting. I spend an hour a few days ago iterating on a document where I was basically just fixing bullets and headers. Most alternatives I've tried aren't any better at this.
None of this seems particularly hard; it's just a lot of integration work that just hasn't happened yet. We have a bunch of lego bricks, not a lot of fully mature solutions. MCP isn't a full solution, it's a pretty lego brick. Mostly even OpenAI and Anthropic aren't getting around to doing much more than simplistic technology demos with their own stuff. IMHO their product teams are a lot less remarkable than their AI teams.
Unless the MCP server itself has an LLM call inside of it (rare), the MCP server is pretty deterministic. It’s the AI that invokes it that’s actually indeterministic, but the user is already using that.
In the real world, where it is (at least in our current state of overall programming language tooling, and the existence of physics) intractable to prove all eventualities and absence of side-effects of executed code, determinism is indeed a spectrum.
If we want to be specific here, I would say the "pretty deterministic" is equal to "as deterministic as your typical non-LLM REST API call", which still spans a big range of determinism.
One thing I like about the current AI craze is that several services start to provide user-oriented services again, like 90s and 00s.
Search engine for AI returning only relevant results (no ads and promotions),
online stores showing only relevant items,
news articles in non-clickbait, clear and concise style.
I'm not confident the current AI craze will be net positive for humanity.
But one possible good outcome could be that
many people prefer simple chat UI to interact with services,
most companies have to adopt them and are forced to provide simple, straight, no-nonsense content instead of what they want to sell,
while LLMs are just commodity so unnamed Chinese companies can provide models as good as the one from the most VC-funded company
so they can't enshittify the UX.
We already have AG-UI [0], which has been implemented by frameworks like Microsoft agent framework, pydantic AI and llamaindex. I guess they'll just have to duplicate functionality.
Meta note: I really detest authors using cryptobro terminology “A dropped” and “B is huge” for anything technical. It immediately smells of script kiddies or their latest reincarnation.
Well the cryptobros moved into AI so not surprising.
I know because we are getting 3-4 spam messages per day on our discord server from various developers with previous experience in crypto now looking for opportunities in the agentic AI development space.
It'll be interesting to see how this goes, but my first impression is that it's actually not where we want to go. One of the cool things about MCP (or even just tool calling) is that the LLM on top of a tool provides a highly flexible and dynamic interface to traditionally static tools.
I love being able to type "make an iptables rule that opens 443" instead of having to dig out the man page and remember how to do that. IMO the next natural extension of this is giving the LLM more capability to generate user interfaces so I can interact with stuff exactly bespoke to my task.
This on the other hand seems the other way round, it's like bolting a static interface onto the LLM, which could defeat the purpose of the LLM interface layer in the first place right?
Giving LLM the ability to generate UI is a cool concept, but our models are not there yet. MCP Apps can be extremely powerful, for example, you can play Doom inside ChatGPT: https://x.com/rauchg/status/1978235161398673553?s=20
I don't think we can generate anywhere close to this kind of UI just yet.
We built https://usefractal.dev/ to make it easier for people to build ChatGPT Apps (they are technically MCP Apps) so I have seen the use cases. Most of these use cases LLM cannot generate the UI on the fly.
That's what we're doing at Hallway (https://hallway.com).
UIs should be fully remix-able and not set by the datasource/SaaS. So we built out a system to allow users to use the standard UI or remix apps as they want. Like Val.town, but with a flexible UX/workspace layer. Come check us out!
> giving the LLM more capability to generate user interfaces
This is not dissimilar to the argument that "MCP needs not exist, just tell llm to run commands and curl". Well, llm can do those, and generate user interfaces. It's just they don't work reliably (maybe ever, depending on how you define "reliable").
I guess as engineers we can do some work and create stopgap solutions or we can all sit and wait for someone else (who? when?) to make AGIs in which everything just magically works, reliably.
MCP has already drastically lost utility already thanks to skills - for most things it is easier to just hand the model a CLI that it can run.
I'd imagine the same thing will happen here: It will prove more flexible to not push the model (and user) towards a UI that may not match what the user is trying to accomplish.
To me this seems like something I categorically don't want unless it is purely advisory.
MCPs as a thin layer over existing APIs has lost utility. Custom MCPs for teams that reduces redundant thinking/token consumption and provides more useful context for the agent and decreases mean time to decision making is where MCPs shine.
Something as simple as correlating a git SHA to a CI build takes 10s of seconds and some number of tokens if Claude is utilizing skills (making API calls to the CI server and GitHub itself). If you have an MCP server that Claude feeds a SHA into and gets back a bespoke, organized payload that adds relevant context to its decision making process (such as a unified view of CI, diffs, et. al), then MCP is a win.
MCP shines as a bespoke context engine and fails as a thin API translation layer, basically. And the beauty/elegance is you can use AI to build these context engines.
In your example, you could achieve a similar outcome with a skill that included a custom command-line tool and a brief description of how to use it.
MCPs are specially well suited for cases that need a permanent instance running alongside the coding agent, for example to handle authentication or some long-lived service that is too cumbersome to launch every time the tool is called.
I mention mean-time to decision making and that's one of the rationales for the mcp. A skill could call a script that does the same thing -- but at that point aren't we just splitting hairs? We are both talking about automated repetitive thinking + actions that the agent takes? And if the skill requires authentication, you have to encode passing that auth into the prompt. MCP servers can just read tokens from the filesystem at call time and don't require thinking at all.
Exactly. The way it’s mostly been used so far is a poor abstraction over stuff you can just put in the context and have the agent run commands.
It really shines in custom implementations coupled to projects. I’ve got a QT desktop app and my mcp server allows the agents to run the app in headless mode, take screenshots, execute code like in Playwright, inspect widget trees, send clicks/text/etc with only six tools and a thousand tokens or so of instructions. Took an hour to build with Claude Code and now it can run acceptance tests before committing them to code end to end tests.
There are important contexts outside of machines you control where installing or running cli commands isn’t possible. In those cases, skills won’t help, but MCP will.
Hence why I said drastically, rather than totally. There are still a few edge cases where it is worthwhile, but they are small and shrinking, especially with services providing UI's with VM's/containers for the model to use increasingly being a thing.
I personally don’t see why developers should just add tons of functionality to any model for free like this. Some of these MCPs are pretty good, and I was a little shocked how much functionality developers released for free to drop into something like Claude. Either developers are stupid or there really is no market yet.
> for free like this
for free: not.
but I'm certain the race has just begun: big service providers and online retailers are currently implementing widgets enabling the purchase of their services and goods directly within the ChatGPT or Claude chat windows.
Disclosure: I work at Anthropic, have worked on MCP
I also think this is pretty big. I think a problem we collectively have right now is that getting MCP closer to real user flows is pretty hard and requires a lot of handholding. Ideally, most users of MCP wouldn't even know that MCP is a thing - the same way your average user of the web has no idea about DNS/HTTP/WebSockets. They just know that the browser helps them look at puppy pictures, connect with friends, or get some work done.
I think this is a meaningful step in the direction of getting more people who'll never know or care about MCP to get value out of MCP.
I want to try and understand what you guys see as the win from MCP. It's objectively inferior to code/clis across a ton of dimensions. The main value I see from it is as a single point to "sandbox" what your agents can do, but it seems a little awkward for that use case.
LLMs are used outside of programming though. It's much easier to hook up a HTTP MCP than it is to install and update a CLI on the execs machines.
I built https://terminalwire.com around the idea that CLIs are a great way to interact with web applications.
Turns out the approach works well for integrating web apps with LLMs. I have a payroll company using it in their stack to replace MCP and they’re reporting lower token usage and a better end result.
Re CLI: In fact it seems to be very similar to command line in AutoCAD (you can do things visually with mouse or choose to draw via CLI). With LLMs it is more sophisticated (intelligent) because you are not limited with set of predefined commands.
I am waiting for Excel CLI…
I want to also highlight that this is an optional "extension" to MCP, not a required part of specification. If you are a terminal application, only care about tool calling, etc, you are free to skip this. If you want to enable rich clients, then it might be something to consider implementing.
Can you potentially review utcp.io for seeing what MCP should have been to avoid context overloading? The dynamic tool search option and codemode to avoid loading all tools/servers into context seems like a better method imo.
I wonder how long it'll take you to figure out that you're trying to reinvent deterministic APIs.
Or just APIs in general.
MCP is incredibly vibe-coded. We know how to make APIs. We know how to make two-way communications. And yet "let's invent new terminology that makes little sense and awkward workarounds on top of unidirectional protocols and call it the best thing since sliced cheese".
MCP is not just providing an API, it’s providing a client that uses that API.
And it does so in a standard way so that my client is available across AI providers. My users can install my client from an ordained URL without getting phished, or the LLM asking them to enter an API key.
What’s the alternative? Providing a sandbox to execute arbitrary code and make API calls? Having an LLM implement OAuth on the fly when it needs to make an API call?
MCP has a place.
> MCP is not just providing an API
It does just provide an API. Your client may have a way to talk to some software via MCP protocol. You know, like a client can talk to a server exposing an endpoint via an API.
> And it does so in a standard way so that my client is available across AI providers.
As in: it's an API on a port with a schema that a certain subset of software understands.
> What’s the alternative? Providing a sandbox to execute arbitrary code and make API calls?
MCP is a protocol. It couldn't care less what you do with your "tool" calls. Vast majority of clients and servers don't run in any sandbox at all. Because MCP is a protocol, not a docker container.
> Having an LLM implement OAuth on the fly when it needs to make an API call?
Yes, MCP has also a bolted-on authorisation that they didn't even think of when they vibe-coded the protocol. And at least finally there was some adult in the room that said "perhaps you should actually use a standardised way to do this". You know, like all other APIs get OAuth (and other types) of authorisations.
Perhaps confusingly, I’m referring to MCP as the sum of the protocol, a server adhering to the protocol, and clients adding support (e.g. “Connectors”).
The combination of these things turns into an ecosystem.
MCP is a Protocol. The server and the clients are just that. It truly is a rebranding of “API” seemingly just because it’s for a specific purpose. Not that there’s anything wrong with that… call it whatever. But I don’t understand the need to sell it as something else entirely. It is quite literally a reinvention of RPC.
> I’m referring to MCP as the sum of the protocol, a server adhering to the protocol, and clients adding support (e.g. “Connectors”).
Why?
Do you refer to REST APIs or GraphQL as a whole? There are servers "adhering to the protocol" and "clients adding support" for these.
These are literally APIs.
What is indeterministic about MCP servers? Most of them follow fairly simple rules, eg an MCP server to interact with Slack gives pretty deterministic responses to requests.
Or are you confusing the LLM / MCP client invoking the tools being non-deterministic?
MCP is already deterministic. What's huge about it is that it has automatic API discovery and integration built-in. It's a bit rough yet but I think we will only see how it's getting improved more and more.
> automatic API discovery and integration
So, WSDL?
Yes, WSDL is great for it, now go ahead and convince people. What is your marketing budget?
You really want to use WSDL? OpenAPI v3 would be a much better fit. But it has a tone of features that are completely unnecessary for this use-case. What if we just stripped it down to input and output json schemas? Oh wait ... we just invented MCP.
WSDL just told you what API was there and create the interface for you in code.
It actually never tried to figure out what API call you actually needed based on what the user asked and how to handle it in real time.
I mean it could, but WSDL was already superseded by REST.
no, thank you
I am a pretty big proponent of MCP, and I think this at least for now is not a move in a good direction.
The whole surface of the MCP specification is already pretty big, and barely any server implements anything beyond the core parts.
With elicitation there was already a lightweight version of this in place in the standard, and I'm not sure I've ever encountered a server or client implementation of it in the wild, and elicitation is an order of magnitude simpler to integrate on a conceptional level.
I fear that this has a significant risk of splintering the MCP ecosystem further (it's already pretty strained due to the transport protocol iterations), and there isn't really a reason to create a official extension (yet), that may worst case also require multiple iterations to get things right.
All it means that there is a common place and a recommendation that if you care as a client implementor (e.g. postman, chatgpt and others do), there is a spot to look for a generally accepted way of how to do that instead of increasing amount of proprietary extensions on top of MCP.
> there is a spot to look for a generally accepted way of how to do that
I think MCP-UI in it's current state can fill that role very well, and that's not the only lens with which to view a SEP like this.
From what I can tell the main thing that this SEP does is put the official blessing of the MCP project on the existing underlying protocol that MCP-UI is already using. I think the better way would be to let MCP-UI and competing implementations cook for a little bit longer, rather than trying to get an officially sanctioned extension out the door and then iterate on that, which will cause a lot more churn.
For a non-trivial feature like this, I would expect an SEP that highlights more alternative usage scenarios and potentially operates on a more technology agonistic layer (e.g. it's very JS+iframe bound, so what about mobile or terminal UIs?), similar to the main MCP spec. Especially given the backdrop of the main MCP specification, which feels much more well-rounded and throughly considered (though in a few areas still incomplete), this SEP does not seem to meet the same bar.
Reading that from the perspective of someone that builds a LLM Chat UI[0] (and aims to be able to maintain that long-term) that heavily builds on MCP as an interoperability concept, reading what is proposed here and seeing how prescriptive it is in many aspects does not spark a lot of joy.
[0]: https://erato.chat
---
EDIT: From what I can tell I'm replying to one of the co-creators of MCP. So let me just ask quite directly: Do you think that the risk of proprietary sprawl in terms of MCP-UI alternatives is currently greater than prematurely standardizing a potentially incomplete version?
I appreciate the take. I can certainly see your position.
In my mind what we are doing is building a common place to evolve the pattern. I wouldn’t really call it “standardize”, since in the end in this space adoption matters. (Writing a standard is worth nothing if nobody is using it). I expect MCP apps to evolve and iterate independent from the main spec for a while. You are right, it’s early and premature to say “this is done”. That’s the goal with extensions.
I do believe that I rather have commonalities between a Claude.ai and a ChatGPT as a developer (not sure that’s true if I was mostly looking at it from a product perspective). I also think you will see chat providers iterate on top of it, and mcp apps is more a common core than the full thing one can use on every platform.
Yes, it is huge to spread the work out on embedded UIs in chat interfaces. But I think the design direction is going, is exactly the same direction of how Google Assistant, Amazon Alexa, or any of the other assistants work, ifkyk.
The MCP community is just reinventing, but yes, improving, what we've done before in the previous generation: Microsoft Bot Framework, Speaktoit aka Google Dialogflow, Siri App Shortcuts / Spotlight.
And interactive UIs in chats go back at least 20 years, maybe not with an AI agent attached...
The next thing that will be reinvented is the memory/tool combination, aka a world model.
IMO this is not where we want to go. The future is you have a system agent that interact with all your apps via API, and websites are real-time so you can see what your agent is doing as you do it it. We don't need more rigging for this future, just better API support.
Trying to create custom agent APIs to embed apps in chat is a very "monopolist frontier lab" thing to try and do.
While they're at it, they might as well check if their answers end with a yes/no question, and, if so, offer a "yes" button so that i can answer yes with a single click.
> If you want a focused comparison next - for example, benchmarks on coding/math, token-cost examples for a typical session, or API usage differences - I can produce a compact table with sources and numbers.
--> can be answered with yes, so please add a yes button. A no button is not needed.
How reliable are the processes which these things run?
I'm processing thousands of files using Copilot, and even 20 at a time, it usually skips a couple, and sometimes, when skipping, it merges the data from one file to the next, not applying anything to the second file, other times it completely applies the data parsed from one file to the second --- not a big deal since I'm reviewing each operation manually, but the only reason the error rate is acceptable is the files are so inconsistent that normal techniques weren't working.
Is there an equivalent to "double-keying" where two different LLMs process the same input and it only moves forward if both match perfectly?
You should probably ask the AI to write a script to do the task. Any procedure that needs to be perfect should be done by writing deterministic code.
The task is renaming scans of checks (with the barcode/bank account info obscured by a pen) --- the checks are of varying sizes, placement is not exact, the placement and formatting of the information on the check is essentially random, and many of them are (poorly) handwritten).
The LLM is working well enough for my needs (and I'm using a locked-down computer which installing/running development environments/scripts on is awkward), and it's a marked improvement over the previous technique of opening 50 files at a time, noting the Invoice ID, closing the file, typing the Invoice ID as a name, then quitting Adobe Acrobat and re-launching it for the next 50 (if that was not done, eventually Acrobat would reach a state where it would close a file and despite the name having been typed, not save it), then using a .bat file made using concatenation in an Excel column.
It would be nice if it were perfect, but each check has to be manually entered, and the filename updated to match the entry by hand.
test how much data from a file they actually see.. its not unlimited// build a deterministic output ie. 2 of 5 chunks received, send chunk 3
Original title: MCP Apps: Extending servers with interactive user interfaces
The post title is quite editorialized.
Correct, the title should be changed.
I skimmed over this, but did I see a reference sandbox implementation? And then basically the chat UI interacts with that with postMessage (and receiving) and forwards tool calls to the MCP server. Does it also forward tool calls the MCP server doesn't handle to the host backend?
What I am imagining is something like a meta UI tool call that just creates a menu. The whole MCP server's purpose might be to add this menu creation capability to the chat user interface. But what you are selecting from isn't known ahead of time, it's the input to the UI.
When they select something I assume it would output a tool call like menuItemSelected('option B'). I suppose if you want your server to do anything specific with this then you would have to handle that in the particular server. But I guess you could also just have a tool call that just sends the inputs to the agent. This could make for what is a very slow to respond but extremely flexible overall UX.
I guess this is not the intended use, but suppose you give your agent generic MCP UI tools for showing any menu, showing any data table, showing a form, etc. So the inputSchemas would be somehow (if this is possible) quite loosely defined.
I guess the purpose is probably more about not having to go through the LLM rather than giving it the ability to dynamically put up UI elements that it has to react to individual interactions with.
But maybe one of the inputs to the dataTable are the query parameters for its data, and the table has a refresh button. Maybe another input is the URI for the details form MCP UI that slides over when you click a row.
Maybe there is an MCP UI for Layout what allows you to embed other MCP UIs in a specific structure.
This might not make sense, but I am wondering if I can use MCP Apps as an alternative to always building custom MindRoot plugins (my Python/web components agentic app framework) to provide unique web pages and UI for each client's agentic application.
I think I may have gotten the MCP Apps and MCP UI a bit conflated here so I probably need to read it again.
I'm not sure I get why we need something specific like MCP-UI? why wouldn't "just another tool" do exactly the same?
Eg you present a "display-graph-chart" tool as a MCP tool, and the agent calls it, it doesn't need to adhere to any protocol except the basic existing MCP protocol, and the UI that's used to interact with the agent would know the best presentation (show it as an embedded HTML graph if in a web ui, show it as a ascii chart if in a terminal, etc)?
Is the idea just to standardize the "output format" of the tool so that any agent UI could display stuff in the same way? so that one tool could work with any agent display?
This is so redundant it is beyond reason.
Given LLMs can generate code complex frontend code, why is so difficult for Antropic / OpenAI to prompt their chat applications to create UI on the fly that matches 100% their Chat applications?
I know this is possible because this is how we do it.
The LLM generates some text that we know how to interpret and we render it on the screen.
Besides, this is exactly how their canvas thing works (both chtgpt and claude) when rendering documents on the side.
Also, this is how GitHub Copilot works when rendering UI to preview draft tickets. It is simply a markdown fenced codeblock with some special properties in the language definition.
Not only this approach is more specific to the application where the UI is supposed to render but it also opens the door for the user to customise how the UI should work by adding their own prompts.
The approach that OpenAI has taken is not better - it is simply a mechanism to create some sort of app store - something they have attempted to do many times in the past as well.
It's very interesting of course at a natural move for aggregation. I'd love to see Stratechery / Ben Thompson's take on this.
From my perspective the challenges for vendors and SaaS providers are [1] discovery [2] monetization [3] disintermediation
I think it's less of a concern if you're Shopify or those large companies that have existing brand moats.
But if you're a startup, I don't think MCP as a channel is a clear-cut decision. Maybe you can get distribution but monetization is not defined.
Also I'm sure the model providers will capture usage data and could easily disintermediate you , especially if your startup is just a narrow set of prompts and a UX over a specific workflow.
The Reforge guys have been talking about a channel shift and this being it but until incentives are clear I'm not sure this is it yet. Maybe an evolution of this.
I'm building an AI coach for job seekers / early stage professionals (Socratify) and while I'd love more distribution from MCP UI integration I think at this point risk is higher than reward...
What’s the difference between an MCP server and a command line executable (which curl is also one)? A good ——help or openapi spec…?
vs executable - don’t need a full shell environment to run it
vs openapi - there’s some more advanced agent-specific concepts in MCP, but primarily, I would say, it’s convention and optimizing for the client. Existing openapi specs will often have not great descriptions, or be a huge mess, or just be huge. Making an MCP server requires you to rethink the UX and tools to be optimized for the access patterns of an agent, and also makes you test it with them.
That said, Claude Skills[0] is something more in the vein of cli with —help (really, with a markdown guide), embracing what is already there with a bit of instructions, as opposed to building from scratch.
My personal opinion, having built multiple MCP servers, is that long-term they will not be the primary approach to tools, and that skills for instance are a better approach for most use-cases. But they do have their use-cases.
[0]: https://www.claude.com/blog/skills
You don't need a shell to run executables either.
MCP defines other useful things like Resources, Elicitation, Sampling, prompts which are model/agent-specific, there is nothing in openai or a "good --help" that could match that.
If you don't see this as useful, that is ok, but it is not the same as a command line executable.
Building MCP Apps (MCP servers with Apps SDK support) is pretty painful right now.
You need to:
1. Spin up a server that returns UI components.
2. Hand-write a bunch of JSON schemas + tool wiring
So we open-sourced a high-level MCP Server SDK that basically lets you have both the MCP server and React components in the same place:
- Every React component you put in your resources/ folder is automatically built and exposed as an MCP resource + tools. No extra registration boilerplate.
- We added a useWidget hook that takes the tool args and maps them directly into your component props, so the agent effectively “knows” what data the widget needs to render. You focus on UI + logic, the SDK handles the plumbing
Docs for that flow here: https://docs.mcp-use.com/typescript/server/creating-apps-sdk...
We also shipped an MCP Inspector to make the dev loop much less painful: you can connect your MCP server, test UI components from tools (with auto-refresh), and debug how it behaves with ChatGPT/agents as you iterate. https://docs.mcp-use.com/inspector/debugging-chatgpt-apps
Both the SDK and the Inspector are open-source, and any contributions are very welcome :)
Those are the repos:
- SDK: https://github.com/mcp-use/mcp-use
- Inspector: https://github.com/mcp-use/mcp-use/tree/main/libraries/types...
I dont think people realize how important this is.
If one of the vendors manages to get their protocol to become the target platform (eg oai and app sdk), that is essentially their vendor lock in to become the next iOS/Android.
Private API’s or EEE strategies are gonna be something to keep an eye for and i wish regulators would step in to prevent them before its too late.
How is it any better if instead of _one_ vendor, _two_ vendors push an immature version of a standards extension that mainly caters to their needs and give it the official stamp of approval under the MCP umbrella?
This is so backwards it's scary.
Having a chatbot that drives websites inside of it is such an attempted monopolist play. Having a system agent that can interact with apps via API without being connected to the app is the pattern that's both elegant and preserves freedom.
I don't see how this is a step down from existing web applications. Should companies building web applications not be opinionated about their user interfaces? When I look at Notion, I should just get any view of the data inside it, regardless of whether it's the same view as my coworker gets? How is this preferable?
> Having a system agent that can interact with apps via API without being connected to the app is the pattern that's both elegant and preserves freedom
Does this prevent anyone from doing that?
Hasn't even Anthropic admitted that MCP is very wasteful of context? Having the LLM write code to call and API can use 98% less tokens.
https://www.anthropic.com/engineering/code-execution-with-mc...
The agent discovers tools by exploring the filesystem: listing the ./servers/ directory to find available servers (like google-drive and salesforce), then reading the specific tool files it needs (like getDocument.ts and updateRecord.ts) to understand each tool's interface. This lets the agent load only the definitions it needs for the current task. This reduces the token usage from 150,000 tokens to 2,000 tokens—a time and cost saving of 98.7%.
The article you link (and general movements in that area, like e.g. discussed here[0]) don't advocate for abandoning MCP, just that masquerading MCP usage as code usage makes it more efficient. In that system MCP still provides a valuable role to act as a tool discovery protocol, a role that would have to be filled one way or another unless you are building a system with a set of tools that is known ahead of time.
[0]: https://news.ycombinator.com/item?id=45917182
In that case discovery could be served through `ls` or AGENTS.md, right? So when should you really use MCP?
It sounds like you are viewing MCP servers through a narrow lens of local coding agents.
I'm mostly working with MCP servers in the context of enterprise/organizational usage for connecting internal data sources and workflows/internal APIs through a standardized interface. There is no equivalent of a "ls or AGENTS.md" there, as there is no file system and no shell in those contexts.
If there is no file system where do the MCP definitions live and how are the agents aware of them? How are your end users aware of these definitions? There has to be a file somewhere.
> where do the MCP definitions
What concretely do you mean be definitions?
If you mean the configuration of the entry points similar to a .mcp.json for Claude Code, they can exist in any form, but ultimately, that only contains the endpoints of MCP servers (the entry points for discovery for each of the MCP servers).
If you mean the definitions of what tools are part of what MCP server (= the meat of what is involved in tool calling), those definitions are part of the MCP server and are only retrieved at runtime.
> how are the agents aware of them
Same as in Claude Code, the agent (= the control loop for the LLM) contacts the list of MCP endpoints, gathers all of the available tools on each MCP server, and then exposes those tools to the LLM (and handles the interactions between LLM and MCP server).
> There has to be a file somewhere.
Not in any way that's meaningful for an `ls` operation.
How is that different then an OpenAPI definition?
It conceptually isn't that far from OpenAPI defintions, but MCP do make certain affordances for making consumptions by LLMs easier.
But it certainly also differs by how they are typically consumed. In what software do you just provide your consuming client just an OpenAPI schema URL to be retrieved at runtime? In my experience, none.
https://github.com/OpenAPITools/openapi-generator
You can also have an LLM generate a client from an OpenAPI spec
I don’t get it. I like MCP as an interface in general, but I don’t understand the use-case they present. For those of you here who like this idea, what is the killer use case?
To me, this looks less like UI interactions and more like the MCP equivalent of maintaining state. You start your program and “click” buttons until you get the desired result, maintaining a constant state between interactions. Isn’t that currently possible if you passed through something like a session-id back to the LLM?
Am I missing something? I’m struggling to see what a UI makes possible that the current workflow does not.
I also generally see/use MCP in terms of remote access to programs through servers. Perhaps that’s where I’m getting lost. Is this exclusively something for local MCP?
This is the first wave of a new generation of Apps which has the potential to replace/enhance the majority of e-commerce websites.
I wouldn't underestimate it - it's the hammer can break up the information silos we've built up around websites/apps.
Why prompt Gertrude(tm) on Ramazon for a specific thing you need, if you can ask ChatGPT to find you said thing along with UIs to purchase it across all e-commerce platforms that agree to play in this market?
because the secret is that the web runs on advertising/targeted recommendations. Brezos(tm) wants you to actively browse Ramazon so he can harvest your data, search patterns etc. Amazon and most sites like that are very not crawl friendly for this reason. Why would Brezos let Saltman get all the juicy preference data?
> For those of you here who like this idea, what is the killer use case?
# 1
- Hey claude, recommend me next book I could read
- sure, what about '1984'? Here is a button to send it directly to your kindle, it will be $4,99.
# 2
- I am looking for the hotel for two in Viena this weekend
- here are some offers and still available (displays grid view generated by mcp ui, handled by booking.com, each item has a button to "book now" directly)
Thank you for this.
So, basically -- it's a method for the MCP Server to display a UI element to the user allowing the user to perform some kind of action. Current, this UI will be an HTML iframe.
I think I got confused by their barchart example. I thought this would have been just as easily supported by the MCP server sending back a PNG/PDF of a barchart for the main interface to display (I also generally think of barcharts as static content). But, the idea of adding specific user interaction elements (buttons) helps to make the concept more clear to me.
If only we as an industry hadn't abandoned REST APIs none of this would be necessary.
We've known for decades that its useful for APIs to be self documented and for responses to use schemas to define the shape of the data.
XML can be verbose and I understand why people preferred JSON for ease use. Had we stuck with REST for the last 20 years we'd be way ahead on that front, though, both in syntax and tooling.
> Had we stuck with REST for the last 20 years we'd be way ahead on that front
We are where we are because there (sadly) hasn't been a reasonable business case to advance REST API documentation beyond the point of badly-documented OpenAPI schemas where the main utility is in generating type-safe API wrappers across different programming languages.
With MCP, there is at least a name to a new movement to build self-describing APIs, as with the advent of LLMs there is now enough of a utility for it. All other pushes into that direction have died out ~10 years ago.
Correct me if I'm wrong, but I thought OpenAPI was itself an attempt to wedge REST concepts back on top of mostly RPC-based APIs. I.e. add documentation and a standard set of schemas on top of JSON APIs (mostly).
I do think the problem is from business concerns, though, and are a clear predictor that MCP will fail. Coming out of the dotcom bubble those left standing wanted to build moats and walls, not APIs that any third party could easily discover and use. The need for REST only shows up when a new player makes a move to effectively gobble up the role of being the gate keeper to the internet. With scale other companies will follow, but begrudgingly and only for a short while.
Sadly not. One of the constraints of REST is that the API be completely self-descriptive. If you need out-of-band information like an API spec, then it's basically not RESTful by definition.
I'm using a REST API to read your comment and submit my response — I didn't need to reference external HN API docs. The interface here is fully self-describing.
I agree, but don't quite follow how your comment fits in here.
In a REST approach, I'd expect an LLM to need some kind of initial entry point, much like needing the initial URL for the HN home page. From there the LLM should be able to parse and discover possible actions, call those actions, and understand the response only by parsing the results and any schemas provided.
To be honest, I think I meant to put that reply on a different comment.
In any case, we're on the same page about REST. OpenAPI was mentioned though and I don't think that's ever actually compatible with that type of usage.
I don't think I've ever encountered an API following OpenAPI that was completely self describing and could be discovered and navigated from a single entry URL. They're always paired with some swagger docs or other external content. That instantly makes them not RESTful.
But you're totally right that my comment didn't fit. Might have meant it for the grandparent comment? Not sure.
LLMs work great with REST so that's always still an option. MCPs have that nice third party plug and play experience but that doesn't mean we all have to build them.
I make tons of little REST APIs for my agents to use and in the AGENTS.md there's just a list of API entry points with descriptions on what they offer. Agents drive them with `curl` and it all works great.
Doesn’t that force you to give the Agent some generic code execution environment, or does everybody already do that anyways?
We never really implemented REST APIs. We had a bunch of REST-ish APIs.
Anyway - the REST movement served it's purpose - it killed SOAP and forced everyone back to simpler HTTP APIs without tons of over-engineered XML layers so it did well.
HTML is an implementation of a REST API, I'm not sure what you mean that we never implemented REST.
SOAP was a pain, XML wasn't doomed though and we didn't need to throw that baby out with the bath water.
When you think about that, without AI, simple automation could do most of what we would need if app and webapp providers would have been to provide interface. Were instead the big fight was to prevent scraping and to get your info in any other way than being forced to use the crappy app. Like for banking.
many banks have api access now, but maybe only for business accounts?
In Europe we have the PSD2 open banking [1] directive which essentially gives developers a standard api access to banking operations on behalf of a user.
It is a requirement for banks to implement this and the FSA will be on your neck if you are not compliant with the standard.
[1] https://finance.ec.europa.eu/regulation-and-supervision/fina...
To access that API you need a license as an authorised third-party provider (TPP), don't you? So in practice it is only for businesses capable of pushing through the bureaucracy.
In my current personal development workflow with Claude Code, I've switched entirely to using CLI tools and scripts over MCP as the experience is much more deterministic and flexible.
A great example is Github, it's a significantly better dev experience having CC call out to the gh cli for actions than trying to invoke the MCP.
I built https://terminalwire.com to make it easier for more web applications to ship CLIs like gh. Are there any web services like GH that you use in your workflow that don’t have decent AI integration that would benefit from a CLI interface?
My fear is that even for consumer apps, MCP (or successor) will outgrow UIs, becoming the only way to interact with an app or certain feature - placing machines wholly in the center instead of humans. Human software should retain the understandability and atomicity of each step taken.
Can we edit the submission title to match what's in the actual article? Not just because I'm annoyed at how transparently Gen Z the HN title is...
MCP-UI is a game changer for conversational commerce! We’re building a decentralized marketplace using MCP-UI and our own Shopify alternative[0].
0. https://marketplace.openship.org
Sounds interesting, but I'm pretty unimpressed with "specification first" approaches. Build a tool and release it. If it resonates the let that be the basis for future standards.
An, the dream, a cross platform App Store you can install apps into any client application that supports MCP, but is open, free and agentic.
It’s basically a “web App Store” and we side step the existing app stores (and their content guidelines, security restrictions and billing requirements) because it’s all done via a mega app (the MCP client).
How could it go wrong?
If only someone had done this before, we wouldnt be stuck in Apples, etc’s walled gardens…
Seriously though; honest question: this is literally circumventing platform requirements to use the platform app stores. How do you imagine this is going to be allowed?
Is ChatGPT really big enough they can pull the “we’re gonna do it, watcha gonna do?” to Apple?
Who’s going to curate this app store so non technical users (the explicitly stated audience) can discover these MCP apps?
It feels like MCP itself; half baked. Overly ambitious. “We’ll figure the details out later”
The apps are LLM agnostic, so all MCP apps will be portable. Economically, this means developers don’t have convince users to pay $20 a month, these users are already paying that. Devs just have to convince users to buy the app on the platform.
I don’t see this being the future state. We’d be talking about a world where any and all apps exist inside of fucking ChatGPT and that just sounds ridiculous.
Looks like OpenAI, Anthropic, and the MCP-UI team actually worked together on a common standard for MCP Apps: https://blog.modelcontextprotocol.io/posts/2025-11-21-mcp-ap...
Honestly, I think the biggest friction for MCP adoption has been how un-userfriendly it is. It’s great for devs, but not the average users. Users don't always want to chat, sometimes they just want to click a button or adjust a slider. This feels like the answer to that problem.
Full disclosure, I'm partial here because of our work at https://usefractal.dev. We were early adopters when MCP first came out, but we always felt like something was missing. We kept wishing for a UI layer on top, and everyone says it's gonna take forever for the industry to adopt, maybe months, maybe years.
I cannot believe the adoption comes so quickly. I think this is gonna be huge. What do you guys think?
You are touching on an important point. Basically OpenAI and others provide a lot of poorly integrated tools and components. You can build nice things with those but you have to deal with a lot of issues and it's a non trivial amount of work that even they aren't doing apparently. Even such a simple thing as triggering an OAuth signin to get access to models is not part of SDKs. Most developer tools require configuring API keys in some file instead. No normal user is ever going to do that,.
Things like ChatGPT are remarkably limited from a UX/UI point of view. The product can do amazing things but the UI is nothing special. The mac version currently has a bug where option+shift+1 opens a chat window but doesn't give it focus. When I do that from vs code it adds the editor window. But it's completely blind to any browser tab on which I do that. I'm sure there are good reasons for all that. But it strikes me a bit as a work in progress that a good product owner would spot.
With apps some of the more powerful capabilities (llms driving UIs directly, doing things in agentic loops, tool and API usage) are going to require much deeper integrations than are currently there. We get hints of what is possible and nice technology demos. But it's still hard to build more complicated workflows around this. Unless you build your own applications.
We've been staring at this from the point of view of automating some highly tedious stuff that we currently do in our company manually. For example, working with chat GPT seems to involve a lot of copy paste and manually doing things that it can't really do by itself. Even something as simple as working on a document it will do alright work on the text but then make a complete mess of the formatting. I spend an hour a few days ago iterating on a document where I was basically just fixing bullets and headers. Most alternatives I've tried aren't any better at this.
None of this seems particularly hard; it's just a lot of integration work that just hasn't happened yet. We have a bunch of lego bricks, not a lot of fully mature solutions. MCP isn't a full solution, it's a pretty lego brick. Mostly even OpenAI and Anthropic aren't getting around to doing much more than simplistic technology demos with their own stuff. IMHO their product teams are a lot less remarkable than their AI teams.
Who wants a button that has indeterministic actions?
Unless the MCP server itself has an LLM call inside of it (rare), the MCP server is pretty deterministic. It’s the AI that invokes it that’s actually indeterministic, but the user is already using that.
This is a very strict definition of MCP. An agent (with LLM call inside) can be an MCP. Event a UI component can be an MCP.
> pretty deterministic
This is an oxymoron.
I meant “pretty” as in, using a search engine is pretty deterministic, any REST API is deterministic.
MCP servers’ tools are literally just function calls. It’s the LLM MCP client that’s not deterministic, not the MCP server.
No it's not.
In the real world, where it is (at least in our current state of overall programming language tooling, and the existence of physics) intractable to prove all eventualities and absence of side-effects of executed code, determinism is indeed a spectrum.
If we want to be specific here, I would say the "pretty deterministic" is equal to "as deterministic as your typical non-LLM REST API call", which still spans a big range of determinism.
Calling it... Vibe clicking
AI fills in a form, and you want to adjust the form before clicking submit. How often do you have to adjust AI's answer vs accepting it as is
The popularity of slot machines suggests there is a market
> button that has indeterministic actions
google.com (1998-present)
[flagged]
Am I missing something or is this essentially same as GPT Apps that have been introduced a while ago and have been discussed 10000 times.
Turns the concept of GPT Apps into an open standard rather than something ChatGPT only.
So we just reinvented web pages from first principles?
I think I have to be missing what is huge here?
One thing I like about the current AI craze is that several services start to provide user-oriented services again, like 90s and 00s. Search engine for AI returning only relevant results (no ads and promotions), online stores showing only relevant items, news articles in non-clickbait, clear and concise style.
I'm not confident the current AI craze will be net positive for humanity. But one possible good outcome could be that many people prefer simple chat UI to interact with services, most companies have to adopt them and are forced to provide simple, straight, no-nonsense content instead of what they want to sell, while LLMs are just commodity so unnamed Chinese companies can provide models as good as the one from the most VC-funded company so they can't enshittify the UX.
We already have AG-UI [0], which has been implemented by frameworks like Microsoft agent framework, pydantic AI and llamaindex. I guess they'll just have to duplicate functionality.
Sigh.
[0] https://docs.ag-ui.com/introduction
AG-UI is compatible with both MCP-UI and other declarative UI specs.
AG-UI is built to support agentic behavior in your application, more similar to the Vercel AI SDK but horizontal and standardized.
MCP-UI helps external apps and APIs pass UI to applications, really well suited for use cases like Shopify stores inside of ChatGPT.
We had AG-UI. I'd say CopilotKit and AssistantUI (YC W25) are now Sherlocked
?
Meta note: I really detest authors using cryptobro terminology “A dropped” and “B is huge” for anything technical. It immediately smells of script kiddies or their latest reincarnation.
Well the cryptobros moved into AI so not surprising.
I know because we are getting 3-4 spam messages per day on our discord server from various developers with previous experience in crypto now looking for opportunities in the agentic AI development space.
Amazing
2010 Facebook apps
2015 WeChat mini program
...
2025 MCP-UI
I'm tired.
[flagged]