"Authentication is tricky and so it was very fair that the designers chose not to include it in the first version of the protocol."
No, it's not fair at all.
You can't add security afterwards like spreading icing on baked cake.
If you forgot to add sugar to the cake batter,
there's not enough buttercream in the world to fix it.
I haven't read a good justification from them on this but it did seem reasonable to me that if the protocol allows auth to be added on top then it's less critical to be in the spec v1 itself.
It's not no security vs security but not standardized vs standardized.
Agree though that's it's not ideal and there will definitely be non zero harm from that decision.
I made a MCP server to access remote storage using any kind of protocols (S3, SFTP, FTP, .... ), and authentication was not as much of a problem that most people would make you think. Is it technically pure and beautiful? No. Does it work? Yes and you can check by yourself by pointing the mcp inspector to "https://demo.filestash.app/sse" and play around
I’ll admit I’m a total dunce with respect to using MCP but given it uses TLS why not just use mutual TLS and validate the client certificate for authentication? You don’t need to use higher level protocols for authentication if your transport protocol supports it already.
I would think authZ is the trickier unhandled part of MCP as I don’t remember any primitives for authorization denial or granularity, but, again, HTTP provides a coarse authZ exchange protocol.
Presumably STDIO depends on the local AuthN/PAM that enabled the user space IO to begin with? I suspect what people are concerned about is the authorization of the LLM to the agent and its data rather than authN. AuthN in stdio would be the user that started the processes, or went through some AuthN chain allowing stdio to forward (I.e., ssh, sudo, etc). Am I missing something?
Right - that’s my point. It’s janky to put AuthN into a high level protocol like MCP when low level protocols do a better job that’s more reliable and thoroughly vetted. You can do full on SAML etc and go hog wild without ever touching MCP while inheriting all the thorough analysis and implementation pressure testing down stack. Doing something unsafe because you don’t know what safe looks like isn’t application protocols fault.
The bigger issue is that we need a good why to authenticate more dynamically. In the context of a desktop app (desktop claude, claude code) where you are running an application locally, for your own purposes it can run as your own user.
But when you are running things in any kind of multi-user/multi-tenant scenario, this is much harder and the protocol doesn't really address this (though also doesn't prevent us from layering something on top). As a dumb (but real) example, I don't want a web-enabled version of of an MCP plugin to have access to my company's google drive and expose that to all our chat users. That would bypass the RBAC we have. Also I don't want to bake that in at the level of the tool calls, as that can be injected. I need some side channel information on the session to have the client and server to manage that.
Ok fair enough - effectively there needs to be an authorization propagation channel that is systematic and bypasses the LLM and probabilistic layers through to each endpoint through the server. The issue isn’t the MCP server and client themselves but that there’s no clear and defined way to pass authZ through to the other side. AuthN should be handled at the interface and the AuthZ propagation is the crucial missing piece. Right?
There is no need to implement a new form of authentication that's specific to the protocol because you already have a myriad of options available with HTTP.
Any form of auth used to secure a web service can be used with MCP.
It's no different than adding authN to a REST API.
Please just read the spec. It just builds on top of JSON-RPC, there's nothing special or inherently new about this protocol.
Exactly, lots of people without any idea on what's going on. I implemented an authenticated MCP that uses the context bearer token and passes it on to the backend. Fully authenticated and authorized.
I think recommendation != being enforceably part of it or being implemented.
There are a lot of great recs in the docs but I wrote this based on what I actually saw in the wild. I definitely don't think it's all on the spec to solve these.
You didn't read the spec. And you didn't mention Stdio. Stdio is secure and that's what is used by Claude Desktop the first MCP client and they don't support SSE yet. Since the start, SSE was set a draft and there was a long discussion over the authentification implementation first round already got defined and second is defined again.
I saw a lot of articles since MCP was buzzing same claims copy & paste. And the post show a lot of confusion for what MCP is and MCP do.
MCP is a dead end for chatbots. Building valuable workflows requires more than tool calling, most importantly, understanding the context of a conversation to adjust the tools dynamically.
MCP tools are completely static. If you have to expose some APIs bypassing MCP, then you don't need MCP in the first place, because you don't have the tools abstracted anymore.
I believe it is possible to build nuanced workflows with well abstracted / reusable / pluggable tools, it's just not as simple as implementing a static discovery / call dispatch layer.
I'm not the person you asked, but it does jam me up that Claude Desktop currently makes you edit a .json file and restart the app to add new MCP servers[1], so that's at least one static thing. I'm sure someone will show up with a github link of awesome-org/mcp-mcp-server
MCP isn't static. It explicitly includes support for dynamically modifying tools, resources, etc via it's client notifications[0]. Sure, context is usually opaque to the server itself (unless you use the sampling feature[1]), but there's nothing preventing MCP clients/hosts from adjusting or filtering tools on their own.
Server notifications are a bad way of implementing semantic retrieval on tools. There when would one update the tools? You can’t “trigger” an event which causes a tool change without some hacky workarounds
A tool change can send a notification to the client, but the client chooses when to actually update the tools. This could take time and the LLM may not be aware of the new tool. I don’t think there is a concept of a “well-behaved” client since MCP is about giving flexibility of implementation.
I wouldn’t call this ergonomic. Alternatively, you could just notify the server when a user message is sent, and allow the server to adjust the tools and resources prior to execution of the agent (this is clearly different from the MCP spec).
On a separate note, what client are you using that supports notification, I haven’t seen one yet?
It's great for general and open ended conversation systems where there is no predefined flow or process to follow. Where you have general capabilities you add to the agent (web search, code execution, web drivers, etc.). But a lot of agentic architecture patterns aren't like that, there you want to closely model the flow, guard rails and constraints of the agent. You can't just take a bunch of MCP services and plug them together to solve every business problem. I think it's a bit unfair to MCP because it doesn't seem to even attempt to solve these problems in the first place. It's not replacing things like LangChain.
Everyone should welcome MCP as an open community-driven standard, because the alternative are fractured, proprietary and vendor-locked protocols. Even if right now MCP is a pretty bad standard over time it's going to improve. I take a bad standard that can evolve with time, over no standard at all.
Right ... the best way to think about MCP right now flipping the switch from every app/agent developer building their own client wrappers around 3rd party APIs to either the API providers building them or community maintained wrappers. But all that gives it is a large toolbox. Like NPM or PyPi or any Apt or any our package maintainer. You still need systems to orchestrate them, secure them, define how to use them, etc ... The next-gen of langchains will be all the much better for having a big toolchest ... but it doesn't negate the need for us to innovate and figure out what that will look like
I could see some cases were the tools are user data specific. You upload some csv and now there are some tools customized for slicing and manipulating that data.
It's totally possible to build tools in way that everything is static but might be less intuitive for some use cases.
Langchain is one of many frameworks that implement the "tool calling" pattern, where LLM applications can request that a tool is run (eg "search the web for X" or "execute this Python code") and see the result of that.
MCP is a standard for describing and exposing tools to LLM applications.
There's a MCP "server" spec for writing code that exposes tool definitions and the ability to execute them, and an MCP "client" spec for building systems that can make use of those servers.
It's a spec for how external info can get into a language model's context window. Part of that is tool calling, but there's also templated input, cancellation, progress tracking, recursive generation sampling, statefulness of tool servers, etc.
Related:
> Tool-Calling - If you’re like me, when you first saw MCP you were wondering “isn’t that just tool-calling?”...
Not everyone uses langchain nor does langchain cover some of the lower level aspects of actually connecting things up. MCP just helps standardize some of those details so any assistant/integration combo is compatible.
Yeah, I don't feel comfortable downloading tiny server programs from random devs on the internet to give my LLM client apps extra "tools". I can LuLu or LittleSnitch regular apps but not these ones.
I played around with MCP this weekend and I agree. I just want to get a users X and then send X to my endpoint so I can do something with it. I don't need any higher level abstraction than that.
I think it's fine if you only need a standalone API or know exactly which APIs to call. But when users ask questions or you're unsure which APIs to use, MCP can solve this issue—and it can process requests based on your previous messages.
I just want to mention something in a chat in 5 seconds instead of preparing the input data, sending it to the API, parsing the output for the answer, and then doing it all again for every subsequent message.
Because you need mainly a bridge between the Function calling schema defined that you expose to the AI model so you can leverage them.
The model need a gateway as API can't be used directly.
MCP core power is the TOOLS and tools need to translate to function calls and that's mainly what MCP do under the hood. Your tool can be an API, but you need this translation layer function call ==> Tool and MCP sits in the middle
It is an API. You can implement all this stuff from scratch with raw requests library in python if you want. Its the idea of a standard around information interchange, specifically geared around agentic experiences like claude code, (tools like aider, previously, that are much worse) - its like a FastAPI web app framework around building stuff that can helps LLMs and VLMs and model wrapped software in ways that can speak over the network.
I have an API, but I built an MCP around my API that makes it easier for something like Claude to use — normally something that's quite tough to do (giving special tools to Claude).
I'm not sure I get it too. I get the idea of a standard api to connect one or more external resources providers to an llm (each exposing tools + state). Then I need one single standard client-side connector to allow the llm to talk to those external resources- basically something to take care of the network calls or other forms of i/o in my local (llm-side) environment. Is that it?
Sounds mostly correct. The standard LLM tool call 'shape' matches the MCP tool call 'shape' very closely. It's really just a simple standard to support connecting a tool to an agent (and by extension an LLM).
There are other aspects, like Resources, Prompts, Roots, and Sampling. These are all relevant to that LLM<->Agent<->Tools/Data integration.
As with all things AI right now, this is a solution to a current problem in a fast moving problem space.
MCP allows LLM clients you don’t control—like Claude, ChatGPT, Cursor, or VSCode—to interact with your API. Without it, you’d need to build your own custom client using the LLM API, which is far more expensive than just using existing clients like ChatGPT or Claude with a $20 subscription and teaching them how to use your tools.
Not all clients support that—currently limited to ChatGPT custom GPT actions. It’s not a standard. Fortunately, Anthropic, Google, and OpenAI have agreed to adopt MCP as a shared protocol to enable models to use tools. This protocol mainly exists to simplify things for those building LLM-powered clients like Claude, ChatGPT, Cursor, etc. If you want an LLM (through API calls) to interact with an your API, you can’t just hand it an API key and expect it to work—you need to build an Agent for that.
You could do that. But then you need to explain to the LLM how to do the work every time you want to use that tool.
And you also run into the risk that the LLM will randomly fail to use the tool "correctly" every time you want to invoke it. (Either because you forgot to add some information or because the API is a bit non-standard.)
All of this extra explaining and duplication is also going to waste tokens in the context and cost you extra money and time since you need to start over every time.
MCP just wraps all of this into a bundle to make it more efficient for the LLM to use. (It also makes it easier to share these tools with other people.)
Or if you prefer it. Consider that the first time you use a new API you can give these instructions to the LLM and have it use your API. Then you tell it "make me an MCP implementation of this" and then you can reuse it easily in the future.
> You could do that. But then you need to explain to the LLM how to do the work every time you want to use that tool
This reeks of a fundamental misunderstanding of computers and LLMs. We have a way to get a description of APIs over http, it's called an open API spec. Just like how MCP retrieves it's tool specs over MCP.
Why would an llm not be able to download an openai spec + key and put it into the context like MCP does with its custom schema?
In some sense that is actually what MCP is. A way to document APIs and describe how to call them, along with some standardized tooling to expose that documentation and make the calls. MCP hit a sweet spot of just enough abstraction to wrap APIs without complicating things. Of course, since they didn't add a bunch of extra stuff ... that leads allowing users to footgun themselves per the article.
Yes, but then you have to add that yourself to every prompt. It would be nice to tell your LLM provider just once "here is a tool you can use" along with a description of the API documentation so that you could use it in a bunch of different chat's without having to remind it every single time. That way, when you want to use the tool you can just ask for the tool without having to provide that detail again and again.
Also, it would be kind of cool if you could tell a desktop LLM client how it could connect to a program running on your machine. It is a similar kind of thing to want to do, but you have to do a different kind of processes exec depending on what OS you are running on. But maybe you just want it to ultimately run a Python script or something like that.
MCP is a meta-api and it basically is that, but with further qualifications that the endpoints and how they work themselves are part of the spec so LLMs can work with them better.
I feel like this is being pushed to get more of the system controlled by the provider's side. After a few years, Anthropic, google, etc might start turning off the api.
Similar to how Google made it very difficult to use IMAP / SMTP with Gmail
I asked myself this question before playing with it a bit. And now I have a slightly better understanding, I think the main reason was created as a way to give access of your local resources (files, envvars, network access...) to your LLM. So it was designed to be something you run locally and the LLM has access.
But there is nothing preventing you making an HTTP call from a MCP server. In fact, we already have some proxy servers for this exact use-case[0][1].
Coordinator of the authorization RFC linked in this post[1].
The protocol is in very, very early stages and there are a lot of things that still need to be figured out. That being said, I can commend Anthropic on being very open to listening to the community and acting on the feedback. The authorization spec RFC, for example, is a coordinated effort between security experts at Microsoft (my employer), Arcade, Hellō, Auth0/Okta, Stytch, Descope, and quite a few others. The folks at Anthropic set the foundation and welcomed others to help build on it. It will mature and get better.
Impressive to see this level of cross-org coordination on something that appears to be maturing at pace (compared to other consortium-style specs/protocol I've seen attempted)
This reminds me of something Adam Smith said in The Wealth of Nations:
"People of the same trade seldom meet together, even for merriment and diversion, but the conversation ends in a conspiracy against the public, or in some contrivance to raise prices."
Ymmv, but I cannot image that this "innovation" will result in a better outcome for the general public.
Yet another piece of rushed technology that's being heralded as "the way" which will most certainly be discarded when the next hype-able acronym comes along.
The only upside to these technologies being shotgun implemented and promoted is that they'll inevitably lead to a failure that can't be pushed under the rug (and will irreversibly damage the credibility of AI usage in business).
To paraphrase the Bjarne Stroustrop quote: There are only two kinds of language model tool calling protocols: the ones people complain about and the ones nobody uses
It appears Anthropic developed this "standard" in a vacuum with no scrutiny or review and it turns out it's riddled with horrific security issues, ignored by those hyping up the "standard" for more VC money.
Reminds me of the micro-services hype, which that helps the big cloud providers more than it helps startups with less money even with some over-doing it and being left with enormous amount of technical debt and complex diagrams costing them millions to run.
If people are posting bad information or bad arguments, it's enough to respond with good information and good arguments. It's in your interests to do this too, because if you make them without swipes, your arguments will be more credible.
We have to draw some line on good faith vs bad faith arguments though. Not understanding the difference between a UI and API is a stretch and purposefully conflating them just to win a semantic argument is not productive.
The problem is that internet readers are far, far too prone to classify others as being in bad faith, so in practice, "drawing the line" usually amounts to a provocation. This bias is so strong that I don't think people can be persuaded to draw that line more accurately.
Moreover, the concept of good faith / bad faith refers to intent, and we can't know for sure what someone's intent was. So the whole idea of assessing someone else's good-faith level is doomed from the start.
Fortunately, there is a strategy that does work pretty well: assume good faith, and reply to bad information with correct information and bad arguments with better arguments. If the conversation stops being productive, then stop replying. Let the other person have the last word, if need be—it's no big deal, and in cases where they're particularly wrong, that last word is usually self-refuting.
Let’s remind ourselves that MCP was announced to the world in November 2024, only 4 short months ago. The RFC is actively being worked on and evolving.
Another article accurately critiquing the short-comings of this terrible "spec".
This is exactly why MCP is hardly a mature standard and was not designed to be secure at all making it acceptable for AI agents to claim to execute commands but could also be stealing your credentials or running a totally different command such or could download malware.
The spec appears to to be designed by 6 month-old vibe-coding developers learning Javascript with zero scrutiny rather than members of the IETF at leading companies with maximum scrutiny.
Next time, Anthropic should consult professionals that have developed mature standards for decades and learn from bad standards such as JWT and Oauth.
The author makes good general points but seems to be overloading MCP's responsibilities imo. My understanding of MCP is that it just provides a ready-made "doorway" for LLMs to enter and interact with externally managed resources. It's a bridge or gateway. So is it really MCP's fault that it:
>makes it easier to accidentally expose sensitive data.
So does the "forward" button on emails. Maybe be more careful about how your system handles sensitive data.
How about:
>MCP allows for more powerful prompt injections.
This just touches on wider topic of only working with trusted service providers that developers should abide by generally.
As for:
>MCP has no concept or controls for costs.
Rate limit and monitor your own usage. You should anyway. It's not the road's job to make you follow the speed limit.
Finally, many of the other issues seem to be more about coming to terms with delegating to AI agents generally. In any case it's the developer's responsibility to manage all these problems within the boundaries they control. No API should have that many responsibilities.
I think the author's point is that the architecture of MCP is fundamentally extremely high trust between not only your agent software and the integrations, but the (n choose 2) relationships between all of them. We're doing the LLM equivalent of loading code directly into our address space and executing it. This isn't a bad thing, dlopen is incredibly powerful with this power, but the problem being solved with MCP just isn't that level of trust.
The real level of trust is on the order OAuth flows where the data provider has a gun sighted on every integration. Unless something about this protocol and it's implementations change I expect every MCP server to start doing side-channel verification like getting an email "hey your LLM is asking to do thing, click the link to approve." Where in this future it severely inhibits the usefulness of agents in the same vein as Apple's "click the notification to run this automation."
Sure, at first, until the users demand a "always allow this ..." kind of prompt and we are back in the same place.
A lot of these issues seem trivial when we consider having a dozen agents running on tens of thousands of tokens of context. You can envision UIs that take these security concerns into account. I think a lot of the UI solutions will break down if we have hundreds of agents each injecting 10k+ tokens into a 1m+ context. The problems we are solving for today won't hold as LLMs continue to increase in size and complexity.
Totally agree, hopefully it's clear closer to the end that I don't actually expect MCP to solve and be responsible for a lot of this. More so MCP creates a lot of surface area for these issues that app developers and users should be aware of.
“Traffic calming uses physical design and other measures to improve safety for motorists, car drivers, pedestrians and cyclists. It has become a tool to combat speeding and other unsafe behaviours of drivers”
why would anyone accept to expose sensitive data so easily with MCP ?
also MCP does not make AI agents more reliable, it just gives them access to more tools, which can decrease reliability in some cases:https://medium.com/thoughts-on-machine-learning/mcp-is-mostl...
People accept lots of risk in order to do things. LLMs offer so much potential that people want to use so they will try, and it is but through experience that we can learn to mitigate any downsides.
Exactly. It is pretty common for APIs to actually signal this too. Headers to show usage limits or rates. Good error codes (429) with actual documentation on backoff timeframes. If you use instrument your service to respect read and respect the signals it gets, everything moves smoother. Backing stuff like that back into the MCP spec or at least having common conventions that are applied on top will be very useful. Similarly for things like tracking data taint, auth, tracing, etc ... Having a good ecosystem makes everything play together much nicer.
Also extending the metaphor, you can make a road that controls where you go and makes sure you don't stray from it (whether by accident or on purpose): it's called rail, and its safety guarantees come with reduced versatility.
Don't blame roads for not being rail, when you came in a car because you need the flexibility that the train can't give you.
Yeah. That's another in a long line of MCP articles and blogposts that's been coming up over the past few weeks, that can be summarized as "breaking news: this knife is sharp and can cut someone if you swing it at people, it can cut you if you hold it the wrong way, and is not a toy suitable for small children".
Well, yes. A knife cuts things, it's literally its only job. It will cut whatever you swing it at, including people and things you didn't intend to - that's the nature of a general-purpose cutting tool, as opposed to e.g. safety razor or plastic scissors for small children, which are much safer, but can only cut few very specific things.
Now, I get it, young developers don't know that knives and remote access to code execution on a local system are both sharp tools and need to be kept out of reach of small children. But it's one thing to remind people that the tool needs to be handled with care; it's another to blame it on the tool design.
Prompt injection is a consequence of the nature of LLMs, you can't eliminate it without degrading capabilities of the model. No, "in-band signaling" isn't the problem - "control vs. data" separation is not a thing in nature, it's designed into systems, and what makes LLMs useful and general is that they don't have it. Much like people, by the way. Remote MCPs as a Service are a bad idea, but that's not the fault of the protocol - it's the problem of giving power to third parties you don't trust. And so on.
There is technical and process security to be added, but that's mostly around MCP, not in it.
Well. To repurpose you knife analogy, they (we?) duct-taped a knife on an erratic, PRNG-controlled roomba and now discover that people are getting their Achilles tendons sliced. Technically, it's all functioning exactly as intended, but: this knife was designed specifically to be attached to such roombas, and apparently nobody stopped to think whether it was such a great idea.
And admonishments of "don't use it when people are around, but if you do, it's those people's fault when they get cut: they should've be more careful and probably wore some protective foot-gear" while technically accurate, miss the bigger problem. That is, that somebody decided to strap a sharp knife to a roomba and then let it whiz around in the space full of people.
Mind you, we have actual woodcutting table saws with built-in safety measures: they instantly stop when they detect contact with human skin. So you absolutely can have safe knives. They just cost more, and I understand that most people value (other) people's health and lives quite cheaply indeed, and so don't bother buying/designing/or even considering such frivolities.
This is a total tangent, but we can't have 100% safe knives because one of the uses for a knife is to cut meat. (Sawstop the company famously uses hot dogs to simulate human fingers in their demos.)
Yes. Also, equally important is the fact that table saws are not knives. The versatility of a knife was the whole point of using it as an example.
--
EDIT: also no, your comment isn't a tangent - it's exactly on point, and a perfect illustration of why knives are a great analogy. A knife in its archetypal form is at the highest point of its generality as a tool. A cutting surface attached to a handle. There is nothing you could change in this that would improve it without making it less versatile. In particular, there is no change you could make that would make a knife safer without making it less general (adding a handle to the blade was the last such change).
No, you[0] can't add a Sawstop-like system to it, because as you[1] point out, it works by detecting meat - specifically, by detecting the blade coming in contact with something more conductive than wood. Such "safer" knife thus can't be made from non-conductive materials (e.g. ceramics), and it can't be used to work with fresh food, fresh wood, in humid conditions, etc.[2]. You've just turned a general-purpose tool into a highly specialized one - but we already have a better version of this, it's the table saw!
Same pattern will apply to any other idea of redesigning knives to make them safer. Add a blade cage of some sort? Been done, plenty of that around your kitchen, none of it will be useful in a workshop. Make knife retractable and add a biometric lock? Now you can't easily share the knife with someone else[3], and you've introduced so many operational problems it isn't even funny.
And so on, and so on; you might think that with enough sensors and a sufficiently smart AI, a perfectly safe knife could be made - but then, that's also exist, it's called you the person who is wielding the knife.
To end this essay my original witty comment has now become, I'll spell it out: like a knife, LLMs are by design general-purpose tools. You can make them increasingly safer by sacrificing some aspects of their functionality. You cannot keep them fully general and make them strictly safer, because the meaning of "safety" is itself highly situational. If you feel the tool is too dangerous for your use case, then don't use it. Use a table saw for cutting wood, use a safety razor for shaving, use a command line and your brain for dealing with untrusted third-party software - or don't, but then don't go around blaming the knife or the LLM when you hurt yourself by choosing to use too powerful a tool for the job at hand. Take responsibility, or stick to Fisher-Price alternatives.
Yes, this is a long-winded way of saying: what's wrong with MCP is that a bunch of companies are now trying to convince you to use it in a dangerous way. Don't. Your carelessness is your loss, but their win. LLMs + local code execution + untrusted third parties don't mix (neither do they mix if you remove "LLMs", but that's another thing people still fail to grasp).
As for solutions to make systems involving LLMs safer and more secure - again, look at how society handles knives, or how we secure organizations in general. The measures are built around the versatile-but-unsafe parts, and they look less technical, and more legal.
(This is to say: one of the major measures we need to introduce is to treat attempts at fooling LLMs the same way as fooling people - up to and including criminalizing them in some scenarios.)
--
[0] - The "generic you".
[1] - 'dharmab
[2] - And then if you use it to cut through wet stuff, the scaled-down protection systems will likely break your wrist; so much for safety.
[3] - Which could easily become a lethal problem in an emergency, or in combat.
The problem with the “knife is sharp” argument is that it’s too generic. It can be deployed against most safety improvements. The modern world is built on driving accident rates down to near-zero. That’s why we have specialized tools like safety razors. Figuring out what to do to reduce accident rates is what postmortems are for - we don’t just blame human error, we try to fix things systematically.
As usual, the question is what counts as a reasonable safety improvement, and to do that we would need to go into the details.
I’m wondering what you think of the CaMeL proposal?
Love the trollishness/carelessness of your post. Exactly as you put it: "it is not the road's job to limit your speed".
Like a bad urban planner building a 6 lane city road with the 25mph limit and standing there wondering why everyone is doing 65mph in that particular stretch. Maybe sending out the police with speed traps and imposing a bunch of fines to "fix" the issue, or put some rouge on that pig, why not.
Some of the other issues are less important than others, but even if you accept “you have to take responsibility for yourself”, let me quote the article:
> As mentioned in my multi-agent systems post, LLM-reliability often negatively correlates with the amount of instructional context it’s provided. This is in stark contrast to most users, who (maybe deceived by AI hype marketing) believe that the answer to most of their problems will be solved by providing more data and integrations. I expect that as the servers get bigger (i.e. more tools) and users integrate more of them, an assistants performance will degrade all while increasing the cost of every single request. Applications may force the user to pick some subset of the total set of integrated tools to get around this.
I will rephrase it in stronger terms.
MCP does not scale.
It cannot scale beyond a certain threshold.
It is Impossible to add an unlimited number of tools to your agents context without negatively impacting the capability of your agent.
This is a fundamental limitation with the entire concept of MCP and needs addressing far more than auth problems, imo.
You will see posts like “MCP used to be good but now…” as people experience the effects of having many MCP servers enabled.
They interfere with each other.
This is fundamentally and utterly different from installing a package in any normal package system, where not interfering is a fundamental property of package management in general.
Thats the problem with MCP.
As an idea it is different to what people trivially expect from it.
I think this can largely be solved with good UI. For example, if an MCP or tool gets executed that you didn't want to get executed, the UI should provide an easy way to turn it off or to edit the description of that tool to make it more clear when it should be used and should not be used by the agent.
Also, in my experience, there is a huge bump in performance and real-world usage abilities as the context grows. So I definitely don't agree about a negative correlation there, however, in some use cases and with the wrong contexts it certainly can be true.
> For example, if an MCP or tool gets executed that you didn't want to get executed, the UI should provide an easy way to turn it off or to edit the description of that tool to make it more clear when it should be used and should not be used by the agent.
Might this become more simply implemented as multiple individual calls, possibly even to different AI services, chained together with regular application software?
If you are saying why have autonomous agents at all and not just workflows, then obviously the answer is that it just depends on the use case. Most of the time workflows that are not autonomous are much better, but not always, and sometimes they will also include autonomous parts in those workflows
I don't think that could be sufficient to solve the problem.
I'm using Gemini with AI Studio and the size of a 1 million token context window is becoming apparent to me. I have a large conversation, multiple paragraphs of text on each side of the conversation, with only 100k tokens or so. Just scrolling through that conversation is a chore where it becomes easier just to ask the LLM what we were talking about earlier rather than try to find it myself.
So if I have several tools, each of them adding 10k+ context to a query, and all of them reasonable tool requests - I still can't verify that it isn't something "you [I] didn't want to get executed" since that is a vague description of the failure states of tools. I'm not going to read the equivalent of a novel for each and every request.
I say this mostly because I think some level of inspectability would be useful for these larger requests. It just becomes impractical at larger and larger context sizes.
> It is Impossible to add an unlimited number of tools to your agents context without negatively impacting the capability of your agent.
Huh?
MCP servers aren't just for agents, they're for any/all _clients_ that can speak MCP. And capabilities provided by a given MCP server are on-demand, they only incur a cost to the client, and only impact the user context, if/when they're invoked.
> they only incur a cost to the client, and only impact the user context, if/when they're invoked.
Look it up. Look up the cross server injection examples.
I guarantee you this is not true.
An MCP server is at it's heart some 'thing' that provides a set of 'tools' that an LLM can invoke.
This is done by adding a 'tool definition'.
A 'tool definition' is content that goes into the LLM prompt.
That's how it works. How do you imagine an LLM can decide to use a tool? It's only possible if the tool definition is in the prompt.
The API may hide this, but I guarantee you this is how it works.
Putting an arbitrary amount of 3rd party content into your prompts has a direct tangible impact on LLM performance (and cost). The more MCP servers you enable the more you pollute your prompt with tool definitions, and, I assure you, the worse the results are as a result.
Just like pouring any large amount of unrelated crap into your system prompt does.
At a small scale, it's ok; but as you scale up, the LLM performance goes down.
I think makers of LLM “chat bots” like the Claude desktop or cursor have a ways to go when it comes to exposing precisely what the LLM is being promoted.
Because yes, for the LLM to find the MCP servers it needs that info on its prompt. And the software is currently hiding how that information is being exposed. Is it prepended to your own message? Does it put it at the start of the entire context? If yes, wouldn’t real-time changes in tool availability invalidate the entire context? So then does it add it to end of the context window instead?
Like nobody really has this dialed in completely. Somebody needs to make a LLM “front end” that is the raw de-tokenized input and output. Don’t even attempt to structure it. Give me the input blob and output blob.
… I dunno. I wish these tools had ways to do more precise context editing. And more visibility. It would help make more informed choices on what to prompt the model with.
/Ramble mode off.
But slightly more serious; what is the token cost for a MCP tool? Like the llm needs its name, a description, parameters… so maybe like 100 tokens max per tool? It’s not a lot but it isn’t nothing either.
> An MCP server is at it's heart some 'thing' that provides a set of 'tools' that an LLM can invoke.
A "tool" is one of several capabilities that a MCP server can provide to its callers. Other capabilities include "prompt" and "resource".
> This is done by adding a 'tool definition'. A 'tool definition' is content that goes into the LLM prompt. That's how it works. How do you imagine an LLM can decide to use a tool? It's only possible if the tool definition is in the prompt.
I think you're using an expansive definition of "prompt" that includes not just the input text as provided by the user -- which is generally what most people understand "prompt" to mean -- but also all available user- and client-specific metadata. That's fine, just want to make it explicit.
With this framing, I agree with you, that every MCP server added to a client -- whether that's Claude.app, or some MyAgent, or whatever -- adds some amount of overhead to that client. But that overhead is gonna be fixed-cost, and paid one-time at e.g. session initialization, not every time per e.g. request/response. So I'm struggling to imagine a situation where those costs are anything other than statistical line noise, compared to the costs of actually processing user requests.
> But that overhead is gonna be fixed-cost, and paid one-time at e.g. session initialization, not every time per e.g. request/response.
I don't work for a foundational model provider, but how do you think the tool definitions get into the LLM? I mean, they aren't fine-tuning a model with your specific tools definitions, right? Your just using OpenAI's base model (or Claude, Gemini, etc.) So at some point the tool definitions have to be added to the prompt. It is just getting added to the prompt auto-magically by the foundation provider. That means it is eating up some context window, just a portion of the context window that is normally reserved for the provider, a section of the final prompt that you don't get to see (or alter).
Again, while I don't work for these companies or implement these features, I cannot fathom how the feature could work unless it was added to every request. And so the original point of the thread author stands.
You're totally right, in that: whatever MCP servers your client is configured to know about, have a set of capabilities, each of which have some kind of definition, all of which need to be provided to the LLM, somehow, in order to be usable.
And you're totally right that the LLM is usually general-purpose, so the MCP details aren't trained or baked-in, and need to be provided by the client. And those details probably gonna eat up some tokens for sure. But they don't necessarily need to be included with every request!
Interactions with LLMs aren't stateless request/response, they're session-based. And you generally send over metadata like what we're discussing here, or user-defined preferences/memory, or etc., as part of session initialization. This stuff isn't really part of a "prompt" at least as that concept is commonly understood.
I think we are confusing the word "prompt" here leading to miscommunication.
There is the prompt that I, as a user, send to OpenAI which then gets used. There there is "prompt" which is being sent to the LLM. I don't know how these things are talked about internally at the company. But they take the "prompt" you send them and add a bunch of extra stuff to it. For example, they add in their own system message and they will add your system message. So you end up with something like <OpenAI system message> + <User system message> + <user prompt>. That creates a "final prompt" that gets sent to the LLM. I'm sure we both agree on that.
With MCP, we are also adding in <tool description> to that final prompt. Again, it seems we are agreed on that.
So the final piece of the argument is, as that "final prompt" (or whatever is the correct term) is growing. It is the size of the provider system prompt, plus the size of the user system prompt, plus the size of the tool description, plus the size of the actual user prompt. You have to pay that "final prompt" cost for each and every request you make.
If the size of the "final prompt" affects the performance of the LLM, such that very large "final prompt" sizes adversely affect performance, than it stands to reason that adding many tool definitions to a request will eventually degrade the LLM performance.
> With MCP, we are also adding in <tool description> to that final prompt. Again, it seems we are agreed on that.
Interactions with a LLM are session-based, when you create a session there is some information sent over _once_ as part of that session construction, that information applies to all interactions made via that session. That initial data includes contextual information, like user preferences, model configuration as specified by your client, and MCP server definitions. When you type some stuff and hit enter that is a user prompt that may get hydrated with some additional stuff before it gets sent out, but it doesn't include any of that initial data stuff provided at the start of the session.
> that information applies to all interactions made via that session
Humm.. maybe you should run an llama.cpp server in debug mode and review the content that goes to the actual LLM; you can do that with the verbose flag or `OLLAMA_DEBUG=1` (if you use ollama).
What you are describing is not how it works.
There is no such thing as an LLM 'session'.
That is a higher level abstraction that sits on top of an API that just means some server is caching part of your prompt and taking some fragment you typed in the UI and combining them on the server side before feeding them to the LLM.
It makes no difference how it is implemented technically.
Fundamentally; any request you make which can invoke tools will be transformed, as some point, into a definition that includes the tool definitions before it is passed to the LLM.
That has a specific, measurable cost on LLM performance as the number of tool definitions go up.
The only solution to that is to limit the number of tools you have enabled; which is entirely possible and reasonable to do, by the way.
My point is that adding more and more and more tools doesn't scale and doesn't work.
It only works when you have a few tools.
If you have 50 MCP servers enabled, your requests are probably degraded.
This matches my understanding too, at least how it works with Open AI. To me, that would explain why there's a 20 or 30 question limit for a conversation, because the necessary context that needs to be sent with each request would necessarily grow larger and larger.
"Sequential thinking" is one that I tried recently because so many people recommend it, and I have never, ever, seen the chatbot actually do anything but write to it. It never follows up any of it's chains of thoughts or refers to it's notes.
Simple: if the choice is getting overwhelming to the LLM, then... divide and conquer - add a tool for choosing tools! Can be as simple as another LLM call, with prompt (ugh, "agent") tasked strictly with selecting a subset of available tools that seem most useful for the task at hand, and returning that to "parent"/"main" "agent".
You kept adding more tools and now the tool-master "agent" is overwhelmed by the amount of choice? Simple! Add more "agents" to organize the tools into categories; you can do that up front and stuff the categorization into a database and now it's a rag. Er, RAG module to select tools.
There are so many ways to do it. Using cheaper models for selection to reduce costs, dynamic classification, prioritizing tools already successfully applied in previous chat rounds (and more "agents" to evaluate if a tool application was successful)...
Point being: just keep adding extra layers of indirection, and you'll be fine.
The problem is that even just having the tools in the context can greatly change the output of the model. So there can be utility in the agent seeing contextually relevant tools (RAG as you mentioned, etc. is better than nothing) and a negative utility in hiding all of them behind a "get_tools" request.
I quite liked this article, actually, and I'm quite an MCP stan. These all seem like legitimate problems that the burgeoning ecosystem needs to figure out. Some of them will be logically solved inside the MCP spec, but I also think some won't.
I wrote an MCP Server (called Codebox[1]) which starts a Docker container with your project code mounted. It works quite well, and I've been using it with LibreChat and vscode. In my experience, Agents save 2x the time (over using an LLM traditionally) and is less typing, but at roughly 3x the cost.
The idea is to make the entire Unix toolset available to the LLM (such as ls, find), along with project specific tooling (such as typescript, linters, treesitter). Basically you can load whatever you want into the container, and let the LLM work on your project inside it. This can be done with a VM as well.
I've found this workflow (agentic, driven through a Chat based interface) to be more effective compared to something like Cursor. Will do a Show HN some time next week.
Any interpreter can run malicious code. Mostly the guidance is: don't run malicious code if you don't want it to run. The problem isn't the interpreter/tool but the entity that's using it. Because that's the thing that you should be (mis)-trusting.
The issue is two fold:
- models aren't quite trustworthy yet.
- people put a lot of trust in them anyway.
This friction always exist with security. It's not a technical problem that can or should be solved on the MCP side.
Part of the solution is indeed going to come from containerization. Give MCP agents access to what they need but not more. And part of it is going to come from some common sense and the tool UX providing better transparency into what is happening. Some of the better examples I've seen of Agentic tools work like you outline.
I don't worry too much about the cost. This stuff is getting useful enough that paying a chunk of what normally would go into somebody's salary actually isn't that bad of a deal. And of course cost will come down. My main worry is actually speed. I seem to spend a lot of time waiting for these tools to do their thing. I'd love this stuff to be a bit zippier.
> Give MCP agents access to what they need but not more.
My view is that you should give them (Agents) a computer, with a complete but minimal Linux installation - as a VM or Containerized. This has given me better results, because now it can say fetch information from the internet, or do whatever it wants (but still in the sandbox). Of course, depending on what you're working on, you might decide that internet access is a bad idea, or that it should just see the working copy, or allow only certain websites.
If you give it access to the internet ... it can basically do anything, exfil all your code, receive malicious instructions. The blast radius (presuming it doesn't get out of your sandbox) is limited to loss of whatever your put in (source code) and theft of resources (running a coinminer, host phishing attacks, etc ...). As you say, you can limit things to trusted websites which helps .. but even then, if you trust, say github, anyone can host malicious instructions. The risk tradeoffs (likelihood of of hitting malicious instruction, vs productivity benefit) might nevertheless be worth it ... not to much targetted maliciousness in wild yet. And just a bit more gaurdrailing and logging can go a long way.
This argument comes up a lot, similar to the "humans lie, too" line of reasoning
The difference in your cited case is that employees are a class of legal person which is subject to the laws of the jurisdiction in which they work, along with any legal contracts they signed as a condition of their employment. So, that's a shitload of words to say "there are consequences" which isn't true of a bunch of matrix multiplications that happen to speak English and know how to invoke RPCs
I've heard the opposite, stories from users that MCP has been a flop. One of my acquaintances works in marketing and is a heavy LLM user. He got swept up in the hype of MCPs and tried a bunch of them out. He was expressing his disappointment and frustration.
It is a drawback of hype and early adoption. I think highly technical users can get value out of it for now and can keep their expectations in line. If I am building my own MCP server and it is a bit flaky, I manage that responsibility myself. If I am wiring up a MCP server that makes some claims of automating some workflow and it doesn't work, then I negatively associate that with the very idea of MCP.
I feel MCP emerged too quickly. It will take some time and further protocol versions/enhancements for it to mature, and address stuff like security. Also it was not vetted by a conventional standards body, but just made public by a company. Lots of hype around it, everyone seems to be talking MCP these days.
1 - Claude Desktop (and some more niche AI chat apps) - you can use MCPs to extend these chat systems today. I use a small number daily.
2 - Code Automation tools - they pretty much all have added MCP. Cursor, Claude Code, Cline, VSCode GH Codepilot, etc ...
3 - Agent/LLM automation frameworks. There are a ton of tools to build agentic apps and many support using MCP to to integrate third party APIs with limited to no boilerplate. And if there are are large libraries of every third party system you can imagine (like npm - but for APIs) then these are going to get used.
Still early days - but tons of real use, at least by the early adopter crowd. It isn't just a spec sitting on a shelf for all the many faults.
These examples seem to be tools built on this tool - which is cool and all, but it's not the equivalent of "a mail order catalogue you can access through your computer", or "a replacement for a travel agent".
What are the applications at the level of Amazon.com, Expedia, or Hacker News?
They are little things like “hey I want to get the LLM to fetch logs from elasticsearch” or “hey I want to create a way for the LLM to query my database in a more structured way”.
And yeah in theory openapi can do it but not nearly as token efficient or user efficient. OpenAPI doesn’t help actually “connect” the LLM to anything, it’s not a tool itself but a spec. To use an OpenAPI compliant server you’d still need to tell the LLM how to authenticate, what the server address is, what tool needs to be used to call out (curl?) and even then you’d still need an affordance for the LLM to even make that call to curl. That “afforance” is exactly what MCP defines. It provides a structured way for the LLM to make tool calls.
I'm sure MCP does have legitimate problems and some might even be covered in this article but the author would benefit by spending more time refining their criticisms.
There's a whole section on how people can do things like analyse a combination of slack messages, and how they might use that information. This is more of an argument suggesting agents are dangerous. You can think MCP is a good spec that lets you create dangerous things but conflating these arguments under "mcp bad" is disingenuous.
Id rather have more details and examples on the problem with the spec itself. "You can use it to do bad things" doesn't cut it. I can use http and ssh to bad things too, so it's more interesting to show how Eve might use MCP to do malicious things to Alice or Bob who are trying to use MCP as intended.
I think what is not to be overlooked in these discussions, is the user base of MCP.
In my experience, many less-technical folks started using MCP, and that makes security issues all the more relevant. This audience often lacks intuition around security best-practices. So it’s definitely important to raise awareness around this.
Nothing the author said is wrong, but I don’t know how much it matters or if it would’ve been better if it handled all this out of the gate. I think if MCP were more complicated no one would’ve adopted it.
Being pretty close to OAuth 1.0 and the group that shaped it I’ve seen how new standards emerge, and I think it’s been so long since new standards mattered that people forgot how they happen.
I was one of the first people to criticize MCP when it launched (my comment on the HN announcement specifically mentioned auth) but I respect the groundswell of support it got, and at the end of the day the standard that matters is the one people follow, even if it isn’t the best.
Your experience with rkt is way different from mine. I would gladly accept "podman is..." or even "nerdctl is..." but I hate rkt so much and was thrilled when it disappeared from my life
BetaMax had better fidelity than VHS, but initially could only hold 60 minutes of A/V while VHS could do 120m (ie, almost a full movie without swapping tapes)
"... MCP tends to crowd the model context with too many options. There doesn’t seem to be a clear way to set priorities or a set of good examples to expose MCP server metadata–so your model API calls will just pack all the stuff an MCP server can do and shove it into the context, which is both wasteful of tokens and leads to erratic behavior from models."
Something that I struggle with in regards to MCP and no-ones really touches on:
We don't even know if this "general" Chatbot paradigm will persist. So far I got the feeling that we end up building specialized apps for specialized use-cases (e.g. cursor or claude code for coding) These then bring their own set of tools (for file writing, reading, running bash commands, etc.) and don't really need MCP (except for very niche use-cases maybe).
I don't really see the point yet where LLMs become so good that I throw my specialized LLM tools out and do everything in one claude desktop window. It simply doesn't work generic enough.
Also... if you end up building something custom, you end up having to reimplement the tool calling again anyways. MCP really is just for the user facing chat agents, which is just one section of AI applications. It's not as generically applicable as implied.
If you have a better idea you could get millions in VC funding to pursue that. Chat is where it's at. It's fine to think that it's not gonna be the be-all and end-all of AI, but it's pretty fundamentally how we interact with each other, so I don't see it going away any time soon. There's also a thing as being too generic. So really they just hit this sweet spot.
Fascinating idea. Though I will note that prompt results aren't usually tables when I ask ChatGPT to ask me questions to help me figure out where I personally stand on, say, hot button political issues.
Interestingly, my criticism is exactly the opposite of yours. I think as LLMs become more and more capable (and crucially multi-modal) we will need the external tools less and less.
For example, why would I want an MCP that can drive Photoshop on my behalf? Like I say to the LLM "remove this person from the photo" and it opens Photoshop, uses the magic wand select tool, etc. That is silly in my mind. I want to say "remove this person" and the LLM sends me a perfect image with the person gone.
I extend that idea for just about any purpose. "Edit this video in such and such a way". "Change this audio in such and such a way". "Update this 3d model in such and such a way". No tool needed at all.
And that will lead to more multi-modal input. Like, if I could "mark up" a document with pen marks, or an image. I want tools that are a bit better than language for directing the attention of the model towards the goals I want them to achieve. Those will be less "I am typing text into a chat interface with bubbles" but the overall conversational approach stays intact.
> We don't even know if this "general" Chatbot paradigm will persist
It won't. These startups are selling the sci-fi robot assistant dream; think Tony Stark or Captain Picard or whatever. Once the novelty wears off nobody is going to pay big bucks for what is essentially just childhood nostalgia.
For everything else you'd want hyperspecialized language manipulation tools.
231 comments
[ 3.6 ms ] story [ 309 ms ] threadNo, it's not fair at all. You can't add security afterwards like spreading icing on baked cake. If you forgot to add sugar to the cake batter, there's not enough buttercream in the world to fix it.
It's not no security vs security but not standardized vs standardized.
Agree though that's it's not ideal and there will definitely be non zero harm from that decision.
I would think authZ is the trickier unhandled part of MCP as I don’t remember any primitives for authorization denial or granularity, but, again, HTTP provides a coarse authZ exchange protocol.
You can use whatever authN/authZ you want for the HTTP transport. It's entirely up to the client and server implementers.
But when you are running things in any kind of multi-user/multi-tenant scenario, this is much harder and the protocol doesn't really address this (though also doesn't prevent us from layering something on top). As a dumb (but real) example, I don't want a web-enabled version of of an MCP plugin to have access to my company's google drive and expose that to all our chat users. That would bypass the RBAC we have. Also I don't want to bake that in at the level of the tool calls, as that can be injected. I need some side channel information on the session to have the client and server to manage that.
There is no need to implement a new form of authentication that's specific to the protocol because you already have a myriad of options available with HTTP.
Any form of auth used to secure a web service can be used with MCP. It's no different than adding authN to a REST API.
Please just read the spec. It just builds on top of JSON-RPC, there's nothing special or inherently new about this protocol.
https://modelcontextprotocol.io/specification/2025-03-26
There are way too many commentators like yourself that have no idea what they are talking about because they couldn't be bothered to RTFM.
There are a lot of great recs in the docs but I wrote this based on what I actually saw in the wild. I definitely don't think it's all on the spec to solve these.
I saw a lot of articles since MCP was buzzing same claims copy & paste. And the post show a lot of confusion for what MCP is and MCP do.
I believe it is possible to build nuanced workflows with well abstracted / reusable / pluggable tools, it's just not as simple as implementing a static discovery / call dispatch layer.
1: https://modelcontextprotocol.info/docs/quickstart/user/#3-re...
[0] https://modelcontextprotocol.io/specification/2025-03-26/ser...
[1] https://modelcontextprotocol.io/specification/2025-03-26/cli...
Tool calls can trigger tool changes. Consider an MCP server exposes a list of accounts and tools to manage resources on those accounts:
1. MCP session starts, only tool exposed to the client is the `select_account` and `list_accounts` tools
2. MCP Client selects an account with `select_account` tool
3. MCP Server updates tools for the session to include `account_tool_a`. This automatically dispatches a listChanged notification
4. MCP Client receives notification and updates tools accordingly
IME this is pretty ergonomic and works well with the spec. But that’s assuming your client is well behaved, which many aren’t
I wouldn’t call this ergonomic. Alternatively, you could just notify the server when a user message is sent, and allow the server to adjust the tools and resources prior to execution of the agent (this is clearly different from the MCP spec).
On a separate note, what client are you using that supports notification, I haven’t seen one yet?
For example, why does a “Google search” tool need to change from context to context?
Everyone should welcome MCP as an open community-driven standard, because the alternative are fractured, proprietary and vendor-locked protocols. Even if right now MCP is a pretty bad standard over time it's going to improve. I take a bad standard that can evolve with time, over no standard at all.
It's totally possible to build tools in way that everything is static but might be less intuitive for some use cases.
This just sounds like a small fine tuned model that knows hundreds of MCP tools and chooses the right ones for the current conversation.
MCP is a standard for describing and exposing tools to LLM applications.
So they're not the same category of thing. Langchain could implement aspects of the MCP specification, in fact it looks like they're doing that already: https://github.com/langchain-ai/langchain-mcp-adapters
Related: > Tool-Calling - If you’re like me, when you first saw MCP you were wondering “isn’t that just tool-calling?”...
Not everyone uses langchain nor does langchain cover some of the lower level aspects of actually connecting things up. MCP just helps standardize some of those details so any assistant/integration combo is compatible.
Edit: +1 simonw above
MCP core power is the TOOLS and tools need to translate to function calls and that's mainly what MCP do under the hood. Your tool can be an API, but you need this translation layer function call ==> Tool and MCP sits in the middle
https://platform.openai.com/docs/guides/function-calling
Basically like Rails-for-Skynet
I'm building this: https://github.com/arthurcolle/fortitude
There are other aspects, like Resources, Prompts, Roots, and Sampling. These are all relevant to that LLM<->Agent<->Tools/Data integration.
As with all things AI right now, this is a solution to a current problem in a fast moving problem space.
I built an MCP server that connects to my FM hardware synthesizer via USB and handles sound design for me: https://github.com/zerubeus/elektron-mcp.
And you also run into the risk that the LLM will randomly fail to use the tool "correctly" every time you want to invoke it. (Either because you forgot to add some information or because the API is a bit non-standard.)
All of this extra explaining and duplication is also going to waste tokens in the context and cost you extra money and time since you need to start over every time.
MCP just wraps all of this into a bundle to make it more efficient for the LLM to use. (It also makes it easier to share these tools with other people.)
Or if you prefer it. Consider that the first time you use a new API you can give these instructions to the LLM and have it use your API. Then you tell it "make me an MCP implementation of this" and then you can reuse it easily in the future.
This reeks of a fundamental misunderstanding of computers and LLMs. We have a way to get a description of APIs over http, it's called an open API spec. Just like how MCP retrieves it's tool specs over MCP.
Why would an llm not be able to download an openai spec + key and put it into the context like MCP does with its custom schema?
NIH syndrome, probably.
Also, it would be kind of cool if you could tell a desktop LLM client how it could connect to a program running on your machine. It is a similar kind of thing to want to do, but you have to do a different kind of processes exec depending on what OS you are running on. But maybe you just want it to ultimately run a Python script or something like that.
MCP addresses those two problems.
Did you meant to write "a HTTP API"?
I asked myself this question before playing with it a bit. And now I have a slightly better understanding, I think the main reason was created as a way to give access of your local resources (files, envvars, network access...) to your LLM. So it was designed to be something you run locally and the LLM has access.
But there is nothing preventing you making an HTTP call from a MCP server. In fact, we already have some proxy servers for this exact use-case[0][1].
[0] - https://github.com/sparfenyuk/mcp-proxy
[1] - https://github.com/adamwattis/mcp-proxy-server
https://norahsakal.com/blog/mcp-vs-api-model-context-protoco...
The protocol is in very, very early stages and there are a lot of things that still need to be figured out. That being said, I can commend Anthropic on being very open to listening to the community and acting on the feedback. The authorization spec RFC, for example, is a coordinated effort between security experts at Microsoft (my employer), Arcade, Hellō, Auth0/Okta, Stytch, Descope, and quite a few others. The folks at Anthropic set the foundation and welcomed others to help build on it. It will mature and get better.
[1]: https://github.com/modelcontextprotocol/modelcontextprotocol...
[1]: https://aaronparecki.com/2025/04/03/15/oauth-for-model-conte...
[2]: https://news.ycombinator.com/item?id=43620496
Really enjoyed the article he wrote, just wanted to promote it some more. I learned of several things that will be useful to me beyond MCP.
Congrats to everyone.
"People of the same trade seldom meet together, even for merriment and diversion, but the conversation ends in a conspiracy against the public, or in some contrivance to raise prices."
Ymmv, but I cannot image that this "innovation" will result in a better outcome for the general public.
The only upside to these technologies being shotgun implemented and promoted is that they'll inevitably lead to a failure that can't be pushed under the rug (and will irreversibly damage the credibility of AI usage in business).
It appears Anthropic developed this "standard" in a vacuum with no scrutiny or review and it turns out it's riddled with horrific security issues, ignored by those hyping up the "standard" for more VC money.
Reminds me of the micro-services hype, which that helps the big cloud providers more than it helps startups with less money even with some over-doing it and being left with enormous amount of technical debt and complex diagrams costing them millions to run.
MCP is not a UI. Seem someone here quite confused about what is MCP.
MCP have no security? Someone don't know that stdio is secure and over SSE/HTTP there was already specs: https://modelcontextprotocol.io/specification/2025-03-26/bas....
MCP can run malicious code? Apply to any app you download. How this is the MCP issue? Happen in vscode extensions. NPM libs. But blame MCP.
MCP transmits unstructured text by design?
This is totally funny. It's the tool that decide what to respond. Annd the dialogue is quite
I start feeling this post is a troll.
I stopped reading and even worth continuing over prompt injection and so on.
I'm sure someone in the comments will say that inter-process communication requires auth (－‸ლ.
If people are posting bad information or bad arguments, it's enough to respond with good information and good arguments. It's in your interests to do this too, because if you make them without swipes, your arguments will be more credible.
Moreover, the concept of good faith / bad faith refers to intent, and we can't know for sure what someone's intent was. So the whole idea of assessing someone else's good-faith level is doomed from the start.
Fortunately, there is a strategy that does work pretty well: assume good faith, and reply to bad information with correct information and bad arguments with better arguments. If the conversation stops being productive, then stop replying. Let the other person have the last word, if need be—it's no big deal, and in cases where they're particularly wrong, that last word is usually self-refuting.
Nobody is saying MCP is the only way to run malicious code, just that like VSCode extensions and NPM install scripts it has that problem.
This is exactly why MCP is hardly a mature standard and was not designed to be secure at all making it acceptable for AI agents to claim to execute commands but could also be stealing your credentials or running a totally different command such or could download malware.
The spec appears to to be designed by 6 month-old vibe-coding developers learning Javascript with zero scrutiny rather than members of the IETF at leading companies with maximum scrutiny.
Next time, Anthropic should consult professionals that have developed mature standards for decades and learn from bad standards such as JWT and Oauth.
>makes it easier to accidentally expose sensitive data.
So does the "forward" button on emails. Maybe be more careful about how your system handles sensitive data. How about:
>MCP allows for more powerful prompt injections.
This just touches on wider topic of only working with trusted service providers that developers should abide by generally. As for:
>MCP has no concept or controls for costs.
Rate limit and monitor your own usage. You should anyway. It's not the road's job to make you follow the speed limit.
Finally, many of the other issues seem to be more about coming to terms with delegating to AI agents generally. In any case it's the developer's responsibility to manage all these problems within the boundaries they control. No API should have that many responsibilities.
The real level of trust is on the order OAuth flows where the data provider has a gun sighted on every integration. Unless something about this protocol and it's implementations change I expect every MCP server to start doing side-channel verification like getting an email "hey your LLM is asking to do thing, click the link to approve." Where in this future it severely inhibits the usefulness of agents in the same vein as Apple's "click the notification to run this automation."
A lot of these issues seem trivial when we consider having a dozen agents running on tens of thousands of tokens of context. You can envision UIs that take these security concerns into account. I think a lot of the UI solutions will break down if we have hundreds of agents each injecting 10k+ tokens into a 1m+ context. The problems we are solving for today won't hold as LLMs continue to increase in size and complexity.
In some sense, urban planners do design roads to make you follow the speed limit. https://en.wikipedia.org/wiki/Traffic_calming:
“Traffic calming uses physical design and other measures to improve safety for motorists, car drivers, pedestrians and cyclists. It has become a tool to combat speeding and other unsafe behaviours of drivers”
A better metaphor is the car, not the road. It is legally required to accurately tell you your speed and require deliberate control to increase it.
Even if you stick to a road; whoever made the road is required to research and clearly post speed limits.
Don't blame roads for not being rail, when you came in a car because you need the flexibility that the train can't give you.
Well, yes. A knife cuts things, it's literally its only job. It will cut whatever you swing it at, including people and things you didn't intend to - that's the nature of a general-purpose cutting tool, as opposed to e.g. safety razor or plastic scissors for small children, which are much safer, but can only cut few very specific things.
Now, I get it, young developers don't know that knives and remote access to code execution on a local system are both sharp tools and need to be kept out of reach of small children. But it's one thing to remind people that the tool needs to be handled with care; it's another to blame it on the tool design.
Prompt injection is a consequence of the nature of LLMs, you can't eliminate it without degrading capabilities of the model. No, "in-band signaling" isn't the problem - "control vs. data" separation is not a thing in nature, it's designed into systems, and what makes LLMs useful and general is that they don't have it. Much like people, by the way. Remote MCPs as a Service are a bad idea, but that's not the fault of the protocol - it's the problem of giving power to third parties you don't trust. And so on.
There is technical and process security to be added, but that's mostly around MCP, not in it.
And admonishments of "don't use it when people are around, but if you do, it's those people's fault when they get cut: they should've be more careful and probably wore some protective foot-gear" while technically accurate, miss the bigger problem. That is, that somebody decided to strap a sharp knife to a roomba and then let it whiz around in the space full of people.
Mind you, we have actual woodcutting table saws with built-in safety measures: they instantly stop when they detect contact with human skin. So you absolutely can have safe knives. They just cost more, and I understand that most people value (other) people's health and lives quite cheaply indeed, and so don't bother buying/designing/or even considering such frivolities.
--
EDIT: also no, your comment isn't a tangent - it's exactly on point, and a perfect illustration of why knives are a great analogy. A knife in its archetypal form is at the highest point of its generality as a tool. A cutting surface attached to a handle. There is nothing you could change in this that would improve it without making it less versatile. In particular, there is no change you could make that would make a knife safer without making it less general (adding a handle to the blade was the last such change).
No, you[0] can't add a Sawstop-like system to it, because as you[1] point out, it works by detecting meat - specifically, by detecting the blade coming in contact with something more conductive than wood. Such "safer" knife thus can't be made from non-conductive materials (e.g. ceramics), and it can't be used to work with fresh food, fresh wood, in humid conditions, etc.[2]. You've just turned a general-purpose tool into a highly specialized one - but we already have a better version of this, it's the table saw!
Same pattern will apply to any other idea of redesigning knives to make them safer. Add a blade cage of some sort? Been done, plenty of that around your kitchen, none of it will be useful in a workshop. Make knife retractable and add a biometric lock? Now you can't easily share the knife with someone else[3], and you've introduced so many operational problems it isn't even funny.
And so on, and so on; you might think that with enough sensors and a sufficiently smart AI, a perfectly safe knife could be made - but then, that's also exist, it's called you the person who is wielding the knife.
To end this essay my original witty comment has now become, I'll spell it out: like a knife, LLMs are by design general-purpose tools. You can make them increasingly safer by sacrificing some aspects of their functionality. You cannot keep them fully general and make them strictly safer, because the meaning of "safety" is itself highly situational. If you feel the tool is too dangerous for your use case, then don't use it. Use a table saw for cutting wood, use a safety razor for shaving, use a command line and your brain for dealing with untrusted third-party software - or don't, but then don't go around blaming the knife or the LLM when you hurt yourself by choosing to use too powerful a tool for the job at hand. Take responsibility, or stick to Fisher-Price alternatives.
Yes, this is a long-winded way of saying: what's wrong with MCP is that a bunch of companies are now trying to convince you to use it in a dangerous way. Don't. Your carelessness is your loss, but their win. LLMs + local code execution + untrusted third parties don't mix (neither do they mix if you remove "LLMs", but that's another thing people still fail to grasp).
As for solutions to make systems involving LLMs safer and more secure - again, look at how society handles knives, or how we secure organizations in general. The measures are built around the versatile-but-unsafe parts, and they look less technical, and more legal.
(This is to say: one of the major measures we need to introduce is to treat attempts at fooling LLMs the same way as fooling people - up to and including criminalizing them in some scenarios.)
--
[0] - The "generic you".
[1] - 'dharmab
[2] - And then if you use it to cut through wet stuff, the scaled-down protection systems will likely break your wrist; so much for safety.
[3] - Which could easily become a lethal problem in an emergency, or in combat.
As usual, the question is what counts as a reasonable safety improvement, and to do that we would need to go into the details.
I’m wondering what you think of the CaMeL proposal?
https://simonwillison.net/2025/Apr/11/camel/#atom-everything
Like a bad urban planner building a 6 lane city road with the 25mph limit and standing there wondering why everyone is doing 65mph in that particular stretch. Maybe sending out the police with speed traps and imposing a bunch of fines to "fix" the issue, or put some rouge on that pig, why not.
Good road design makes it impossible to speed.
> As mentioned in my multi-agent systems post, LLM-reliability often negatively correlates with the amount of instructional context it’s provided. This is in stark contrast to most users, who (maybe deceived by AI hype marketing) believe that the answer to most of their problems will be solved by providing more data and integrations. I expect that as the servers get bigger (i.e. more tools) and users integrate more of them, an assistants performance will degrade all while increasing the cost of every single request. Applications may force the user to pick some subset of the total set of integrated tools to get around this.
I will rephrase it in stronger terms.
MCP does not scale.
It cannot scale beyond a certain threshold.
It is Impossible to add an unlimited number of tools to your agents context without negatively impacting the capability of your agent.
This is a fundamental limitation with the entire concept of MCP and needs addressing far more than auth problems, imo.
You will see posts like “MCP used to be good but now…” as people experience the effects of having many MCP servers enabled.
They interfere with each other.
This is fundamentally and utterly different from installing a package in any normal package system, where not interfering is a fundamental property of package management in general.
Thats the problem with MCP.
As an idea it is different to what people trivially expect from it.
Also, in my experience, there is a huge bump in performance and real-world usage abilities as the context grows. So I definitely don't agree about a negative correlation there, however, in some use cases and with the wrong contexts it certainly can be true.
Might this become more simply implemented as multiple individual calls, possibly even to different AI services, chained together with regular application software?
If you are saying why have autonomous agents at all and not just workflows, then obviously the answer is that it just depends on the use case. Most of the time workflows that are not autonomous are much better, but not always, and sometimes they will also include autonomous parts in those workflows
I'm using Gemini with AI Studio and the size of a 1 million token context window is becoming apparent to me. I have a large conversation, multiple paragraphs of text on each side of the conversation, with only 100k tokens or so. Just scrolling through that conversation is a chore where it becomes easier just to ask the LLM what we were talking about earlier rather than try to find it myself.
So if I have several tools, each of them adding 10k+ context to a query, and all of them reasonable tool requests - I still can't verify that it isn't something "you [I] didn't want to get executed" since that is a vague description of the failure states of tools. I'm not going to read the equivalent of a novel for each and every request.
I say this mostly because I think some level of inspectability would be useful for these larger requests. It just becomes impractical at larger and larger context sizes.
Huh?
MCP servers aren't just for agents, they're for any/all _clients_ that can speak MCP. And capabilities provided by a given MCP server are on-demand, they only incur a cost to the client, and only impact the user context, if/when they're invoked.
Look it up. Look up the cross server injection examples.
I guarantee you this is not true.
An MCP server is at it's heart some 'thing' that provides a set of 'tools' that an LLM can invoke.
This is done by adding a 'tool definition'.
A 'tool definition' is content that goes into the LLM prompt.
That's how it works. How do you imagine an LLM can decide to use a tool? It's only possible if the tool definition is in the prompt.
The API may hide this, but I guarantee you this is how it works.
Putting an arbitrary amount of 3rd party content into your prompts has a direct tangible impact on LLM performance (and cost). The more MCP servers you enable the more you pollute your prompt with tool definitions, and, I assure you, the worse the results are as a result.
Just like pouring any large amount of unrelated crap into your system prompt does.
At a small scale, it's ok; but as you scale up, the LLM performance goes down.
Here's some background reading for you:
https://github.com/invariantlabs-ai/mcp-injection-experiment...
https://docs.anthropic.com/en/docs/build-with-claude/tool-us...
Because yes, for the LLM to find the MCP servers it needs that info on its prompt. And the software is currently hiding how that information is being exposed. Is it prepended to your own message? Does it put it at the start of the entire context? If yes, wouldn’t real-time changes in tool availability invalidate the entire context? So then does it add it to end of the context window instead?
Like nobody really has this dialed in completely. Somebody needs to make a LLM “front end” that is the raw de-tokenized input and output. Don’t even attempt to structure it. Give me the input blob and output blob.
… I dunno. I wish these tools had ways to do more precise context editing. And more visibility. It would help make more informed choices on what to prompt the model with.
/Ramble mode off.
But slightly more serious; what is the token cost for a MCP tool? Like the llm needs its name, a description, parameters… so maybe like 100 tokens max per tool? It’s not a lot but it isn’t nothing either.
> An MCP server is at it's heart some 'thing' that provides a set of 'tools' that an LLM can invoke.
A "tool" is one of several capabilities that a MCP server can provide to its callers. Other capabilities include "prompt" and "resource".
> This is done by adding a 'tool definition'. A 'tool definition' is content that goes into the LLM prompt. That's how it works. How do you imagine an LLM can decide to use a tool? It's only possible if the tool definition is in the prompt.
I think you're using an expansive definition of "prompt" that includes not just the input text as provided by the user -- which is generally what most people understand "prompt" to mean -- but also all available user- and client-specific metadata. That's fine, just want to make it explicit.
With this framing, I agree with you, that every MCP server added to a client -- whether that's Claude.app, or some MyAgent, or whatever -- adds some amount of overhead to that client. But that overhead is gonna be fixed-cost, and paid one-time at e.g. session initialization, not every time per e.g. request/response. So I'm struggling to imagine a situation where those costs are anything other than statistical line noise, compared to the costs of actually processing user requests.
> https://docs.anthropic.com/en/docs/build-with-claude/tool-us...
To be clear, this concept of "tool" is completely unrelated to MCP.
> https://github.com/invariantlabs-ai/mcp-injection-experiment...
I don't really understand this repo or its criticisms. The authors wrote a related blog post https://invariantlabs.ai/blog/whatsapp-mcp-exploited which says (among other things) that
> In this blog post, we will demonstrate how an untrusted MCP server ...
But there is no such thing as "an untrusted MCP server". Every MCP server is assumed to be trusted, at least as the protocol is defined today.
I don't work for a foundational model provider, but how do you think the tool definitions get into the LLM? I mean, they aren't fine-tuning a model with your specific tools definitions, right? Your just using OpenAI's base model (or Claude, Gemini, etc.) So at some point the tool definitions have to be added to the prompt. It is just getting added to the prompt auto-magically by the foundation provider. That means it is eating up some context window, just a portion of the context window that is normally reserved for the provider, a section of the final prompt that you don't get to see (or alter).
Again, while I don't work for these companies or implement these features, I cannot fathom how the feature could work unless it was added to every request. And so the original point of the thread author stands.
And you're totally right that the LLM is usually general-purpose, so the MCP details aren't trained or baked-in, and need to be provided by the client. And those details probably gonna eat up some tokens for sure. But they don't necessarily need to be included with every request!
Interactions with LLMs aren't stateless request/response, they're session-based. And you generally send over metadata like what we're discussing here, or user-defined preferences/memory, or etc., as part of session initialization. This stuff isn't really part of a "prompt" at least as that concept is commonly understood.
There is the prompt that I, as a user, send to OpenAI which then gets used. There there is "prompt" which is being sent to the LLM. I don't know how these things are talked about internally at the company. But they take the "prompt" you send them and add a bunch of extra stuff to it. For example, they add in their own system message and they will add your system message. So you end up with something like <OpenAI system message> + <User system message> + <user prompt>. That creates a "final prompt" that gets sent to the LLM. I'm sure we both agree on that.
With MCP, we are also adding in <tool description> to that final prompt. Again, it seems we are agreed on that.
So the final piece of the argument is, as that "final prompt" (or whatever is the correct term) is growing. It is the size of the provider system prompt, plus the size of the user system prompt, plus the size of the tool description, plus the size of the actual user prompt. You have to pay that "final prompt" cost for each and every request you make.
If the size of the "final prompt" affects the performance of the LLM, such that very large "final prompt" sizes adversely affect performance, than it stands to reason that adding many tool definitions to a request will eventually degrade the LLM performance.
Interactions with a LLM are session-based, when you create a session there is some information sent over _once_ as part of that session construction, that information applies to all interactions made via that session. That initial data includes contextual information, like user preferences, model configuration as specified by your client, and MCP server definitions. When you type some stuff and hit enter that is a user prompt that may get hydrated with some additional stuff before it gets sent out, but it doesn't include any of that initial data stuff provided at the start of the session.
Humm.. maybe you should run an llama.cpp server in debug mode and review the content that goes to the actual LLM; you can do that with the verbose flag or `OLLAMA_DEBUG=1` (if you use ollama).
What you are describing is not how it works.
There is no such thing as an LLM 'session'.
That is a higher level abstraction that sits on top of an API that just means some server is caching part of your prompt and taking some fragment you typed in the UI and combining them on the server side before feeding them to the LLM.
It makes no difference how it is implemented technically.
Fundamentally; any request you make which can invoke tools will be transformed, as some point, into a definition that includes the tool definitions before it is passed to the LLM.
That has a specific, measurable cost on LLM performance as the number of tool definitions go up.
The only solution to that is to limit the number of tools you have enabled; which is entirely possible and reasonable to do, by the way.
My point is that adding more and more and more tools doesn't scale and doesn't work.
It only works when you have a few tools.
If you have 50 MCP servers enabled, your requests are probably degraded.
This matches my understanding too, at least how it works with Open AI. To me, that would explain why there's a 20 or 30 question limit for a conversation, because the necessary context that needs to be sent with each request would necessarily grow larger and larger.
I use it in Claude Desktop for the right use case, it's much better than thinking mode.
But, I admit, I haven't tried it in Cursor or with other LLMs yet.
You kept adding more tools and now the tool-master "agent" is overwhelmed by the amount of choice? Simple! Add more "agents" to organize the tools into categories; you can do that up front and stuff the categorization into a database and now it's a rag. Er, RAG module to select tools.
There are so many ways to do it. Using cheaper models for selection to reduce costs, dynamic classification, prioritizing tools already successfully applied in previous chat rounds (and more "agents" to evaluate if a tool application was successful)...
Point being: just keep adding extra layers of indirection, and you'll be fine.
I wrote an MCP Server (called Codebox[1]) which starts a Docker container with your project code mounted. It works quite well, and I've been using it with LibreChat and vscode. In my experience, Agents save 2x the time (over using an LLM traditionally) and is less typing, but at roughly 3x the cost.
The idea is to make the entire Unix toolset available to the LLM (such as ls, find), along with project specific tooling (such as typescript, linters, treesitter). Basically you can load whatever you want into the container, and let the LLM work on your project inside it. This can be done with a VM as well.
I've found this workflow (agentic, driven through a Chat based interface) to be more effective compared to something like Cursor. Will do a Show HN some time next week.
[1]: https://github.com/codespin-ai/codebox-js
The issue is two fold:
- models aren't quite trustworthy yet.
- people put a lot of trust in them anyway.
This friction always exist with security. It's not a technical problem that can or should be solved on the MCP side.
Part of the solution is indeed going to come from containerization. Give MCP agents access to what they need but not more. And part of it is going to come from some common sense and the tool UX providing better transparency into what is happening. Some of the better examples I've seen of Agentic tools work like you outline.
I don't worry too much about the cost. This stuff is getting useful enough that paying a chunk of what normally would go into somebody's salary actually isn't that bad of a deal. And of course cost will come down. My main worry is actually speed. I seem to spend a lot of time waiting for these tools to do their thing. I'd love this stuff to be a bit zippier.
My view is that you should give them (Agents) a computer, with a complete but minimal Linux installation - as a VM or Containerized. This has given me better results, because now it can say fetch information from the internet, or do whatever it wants (but still in the sandbox). Of course, depending on what you're working on, you might decide that internet access is a bad idea, or that it should just see the working copy, or allow only certain websites.
If it is talking to the internet, it is most definitely not sandboxed.
- employees are not necessarily trustworthy
- employers place a lot of trust in them anyway
The difference in your cited case is that employees are a class of legal person which is subject to the laws of the jurisdiction in which they work, along with any legal contracts they signed as a condition of their employment. So, that's a shitload of words to say "there are consequences" which isn't true of a bunch of matrix multiplications that happen to speak English and know how to invoke RPCs
I feel like I hear very many stories of some company integrating with MCP, many fewer stories from users about how it helps them.
It is a drawback of hype and early adoption. I think highly technical users can get value out of it for now and can keep their expectations in line. If I am building my own MCP server and it is a bit flaky, I manage that responsibility myself. If I am wiring up a MCP server that makes some claims of automating some workflow and it doesn't work, then I negatively associate that with the very idea of MCP.
Normally we have a standard when we have applications, but I am not seeing these yet... perhaps I am blind and mad!
1 - Claude Desktop (and some more niche AI chat apps) - you can use MCPs to extend these chat systems today. I use a small number daily.
2 - Code Automation tools - they pretty much all have added MCP. Cursor, Claude Code, Cline, VSCode GH Codepilot, etc ...
3 - Agent/LLM automation frameworks. There are a ton of tools to build agentic apps and many support using MCP to to integrate third party APIs with limited to no boilerplate. And if there are are large libraries of every third party system you can imagine (like npm - but for APIs) then these are going to get used.
Still early days - but tons of real use, at least by the early adopter crowd. It isn't just a spec sitting on a shelf for all the many faults.
What are the applications at the level of Amazon.com, Expedia, or Hacker News?
And yeah in theory openapi can do it but not nearly as token efficient or user efficient. OpenAPI doesn’t help actually “connect” the LLM to anything, it’s not a tool itself but a spec. To use an OpenAPI compliant server you’d still need to tell the LLM how to authenticate, what the server address is, what tool needs to be used to call out (curl?) and even then you’d still need an affordance for the LLM to even make that call to curl. That “afforance” is exactly what MCP defines. It provides a structured way for the LLM to make tool calls.
> We are so back
There's a whole section on how people can do things like analyse a combination of slack messages, and how they might use that information. This is more of an argument suggesting agents are dangerous. You can think MCP is a good spec that lets you create dangerous things but conflating these arguments under "mcp bad" is disingenuous.
Id rather have more details and examples on the problem with the spec itself. "You can use it to do bad things" doesn't cut it. I can use http and ssh to bad things too, so it's more interesting to show how Eve might use MCP to do malicious things to Alice or Bob who are trying to use MCP as intended.
In my experience, many less-technical folks started using MCP, and that makes security issues all the more relevant. This audience often lacks intuition around security best-practices. So it’s definitely important to raise awareness around this.
Being pretty close to OAuth 1.0 and the group that shaped it I’ve seen how new standards emerge, and I think it’s been so long since new standards mattered that people forgot how they happen.
I was one of the first people to criticize MCP when it launched (my comment on the HN announcement specifically mentioned auth) but I respect the groundswell of support it got, and at the end of the day the standard that matters is the one people follow, even if it isn’t the best.
Rkt is better than Docker, later won.
${TBD} is better than MCP, my bet is on MCP.
Your experience with rkt is way different from mine. I would gladly accept "podman is..." or even "nerdctl is..." but I hate rkt so much and was thrilled when it disappeared from my life
"... MCP tends to crowd the model context with too many options. There doesn’t seem to be a clear way to set priorities or a set of good examples to expose MCP server metadata–so your model API calls will just pack all the stuff an MCP server can do and shove it into the context, which is both wasteful of tokens and leads to erratic behavior from models."
I don't really see the point yet where LLMs become so good that I throw my specialized LLM tools out and do everything in one claude desktop window. It simply doesn't work generic enough.
Also... if you end up building something custom, you end up having to reimplement the tool calling again anyways. MCP really is just for the user facing chat agents, which is just one section of AI applications. It's not as generically applicable as implied.
I would like to see:
- Some Smalltalk-like IWE (Integrated Work Environment), creating prompt snippets and chaining them together.
- A spreadsheet like environment. Prompt's result are always tables and you have the usual cell reference available.
For example, why would I want an MCP that can drive Photoshop on my behalf? Like I say to the LLM "remove this person from the photo" and it opens Photoshop, uses the magic wand select tool, etc. That is silly in my mind. I want to say "remove this person" and the LLM sends me a perfect image with the person gone.
I extend that idea for just about any purpose. "Edit this video in such and such a way". "Change this audio in such and such a way". "Update this 3d model in such and such a way". No tool needed at all.
And that will lead to more multi-modal input. Like, if I could "mark up" a document with pen marks, or an image. I want tools that are a bit better than language for directing the attention of the model towards the goals I want them to achieve. Those will be less "I am typing text into a chat interface with bubbles" but the overall conversational approach stays intact.
It won't. These startups are selling the sci-fi robot assistant dream; think Tony Stark or Captain Picard or whatever. Once the novelty wears off nobody is going to pay big bucks for what is essentially just childhood nostalgia.
For everything else you'd want hyperspecialized language manipulation tools.