I run the team at OpenAI that's responsible for the ChatGPT App Store, Codex plugins, and all things MCP.
The thing that all these "MCP is dead" posts are missing is that whether or not MCP is used as a transport protocol is actually completely irrelevant.
The reason MCP isn't dead is because practically ~every company on the planet is building an MCP server. I know this because we interact with all of them. Most of these companies don't have a CLI. Many of these companies don't even have an external API! And yet, they're all building MCP servers.
And that's why MCP is not only not dead, but more important than ever.
Maybe we will turn every MCP server into a CLI under the hood. Maybe we'll use code mode. Maybe we'll implement tool search.
All of those are just implementation details to the much more important point: our AI agents are getting access to services they otherwise would never have had access to.[0] That's what matters.
So, is MCP dead as a direct communication layer for models to speak to? Maybe, maybe not. Is MCP dead as a protocol? Hell no, couldn't be further from the truth.
[0]: Although I will say the Codex app's computer & browser use features have made this statement a lot weaker than it used to be. If you haven't tried them yet—they're mindblowing.
You failed to describe what value the MCP protocol provides.
If all of these companies spent equivalent time writing a CLI for agents to consume as they spend on MCP servers, would they be any worse off in terms of agents being able to interact with their products?
I agree. Mcp might be useless in a personal scenario but it absolutely plays a role of service infrastructure in organizations. It is another form of api for those abilities that are not wrapped with rest api yet. But when they are wrapped in mcp, it seems not necessary to wrap them into rest api or cli again in near future. So these mcp services survive.
The only thing matters is how to import these mcp services into agent context on demand or say by the gradual disclosure principle.
It's not 'who is building' but 'who is using' that's the concern.
AI is a bandwagon tech, a lot of people will 'build because others are' adhering to an ostensible standard.
Most of the people that I know are moving away from MCP in favour of skills where the advantage of MCP goes away if the REST API is clear enough.
Also - I'm sorry to say but MCP management on Codex (and Claude) is just really bad. Everything from discovery, to management, to context window, to documentation - it feels unfinished as a 'feature' even if the protocol is supposed to be narrow.
1) I have a big popup and yellow warning every time a window is opened or a sub agent is launched warning me that 'SkySomething Computer Use' does not work. I had to Google to find out that has something to do with Codex MCP. So already the externalizations of problems, resolutions ... not very well done.
I'm not even sure what to do - and I'm honestly not interested in 'fixing' something I didn't cause, I'm not sure of, and don't want to deal with.
2) Just listing the current MCPs, knowing really what the are for (clearly, concisely) is hard.
What's that supposed to mean? What is 'codex_apps'?
As presented - it resolves to 'nonsense gibberish'. Those are things that I did not even install.
Do you expect people magically know what 'codex_apps' is?
Here is what 'AGI!' Codex 5.5 answered when I asked about 'codex_apps' is:
====
" codex_apps appears to be Codex’s own internal cache/tooling area, not part of J1 (my project).
"I only found it under .codex, e.g.:"
" I did not find it referenced by the J1 source. So unless you saw it somewhere specific, treat it as Codex runtime metadata for app/tool integration, not project code."
====
So even Codex itself has no idea what it's own MCP tools are, and after a full '1 minute of thinking' on 'xhigh' it responded with nonsense.
This whole experience fundamentally deflates my perception of AU, OpenAI, Codex and MCP.
This is supposed to be the 'future' but it feels like 1982 dialup.
This is where 'traditional UX' really starts to show it's value obviously, but you really need to consider enhancing this experience, possibly with some traditional ux mechanisms.
3) Knowing the 'state' of the MCP is totally opaque. Is the 'MCP server' running? Can I restart it? That might be outside the scope of 'Codex' but you're offering the product so all of the underlying stuff is essentially 'your responsibility' as well at least from a UX perspective. Why isn't the 'state' of the MCP listed.
4) How can I not just easily enable/disable individual MCPs so they don't chew up context?
5) How can I not discover MCPs from codex itself, so that I can find solutions to problems? MCPs are all a bit different, and awkward to install and manage. Like with VS Code, we can 'discover plugins'. Even from the Web we can search and discover plugins.
While I realize that most of this rant is oriented around MCP tooling management, and not the standard, I do feel that these issues are 'fundamentally at the crux' of the situation.
Our team has moved away from MCP into Skills - and after doing so, it's hard to see why MCP is going to be valuable - other than plausibly as defining some 'jon calling conventions'.
There's a lot of obvious things to improve, please do that.
This is not even 'basic product design' - it's just 'product common sense'.
That the 'smartest people in the world' have $100 Billion and are are totally scattered on so many issues completely blows my mind but it speaks to how systems are organized.
They don't need to hire anyone for this stuff, they need to have some basic product discipline and prioritize it, that's all.
If they don't do that, all the money in the world and all the best product people wouldn't be able to help.
I totally respect that 'Codex is young' - but it's been kind of a year now. That's a long time - and AI is supposed to 'accelerate time scales' and make people 'super productive' remember?
MCP is essentially just JSON RPC with a few special fields that must be included. I have reservations about JSON RPC, but there needs to be some 'service discovery' layer for LLMs to interface with.
It needs to be available in places like websites, desktop applications, backend services, etc. The CLI is only one place that these systems interface with.
Whatever you replace MCP with will be in a similar shape even if you specify a different communication protocol or different fields for tool discovery.
It's the way that it occupies the context relatively permanently, that it doesn't come along with nice install/uninstall or discovery etc. is the problem.
'Skills' should all be based on MCP, they should load on demand, be very easily manageable and discoverable by humans and by AI, and then it would work
The scope was too narrow, given how it ended up being applied.
If they layer something on top of it, it may yet be revived.
I'll kick myself for not remembering, but there was a fantastic article which suggested that MCP works at org level when unified, safe, access to internal utility APIs need to be given to non-technical staff who do use internal agent tools. Codify your workflow(s) via skills and share across instances, anything that needs context aware API access should be mcp...
hadn't heard of runlayer, but it does make sense. I'm a huge advocate of skills based on the company/process or project owners perspectives and workflow habits rather than using skills.sh or similar. You will end up cosplaying as someone elses perspective and wonder why you don't understand it..
Yes and no -- you can give internal agents access to internal APIs by using rudimentary env var, and org level agentic services tend to offer that kind of permission based access (either roll your own, use an 'enterprise' service, or be knowledgeable that if things go wrong, they'll go very wrong). APIs should, at least from my perspective, always have permission mechanisms. But internal APIs, used by 'internal' agents, have access to those the same way users on the network do, just depends on what flavour of network one is using.
Essentially it's anything that _could_ be on a dashboard, but _might_ be accessed conversationally via an agent.
The article has no date on it, but says deferred tool loading is a recent update that occurred after the article was written. Deferred tool loading was added in Nov 2025: https://www.anthropic.com/engineering/advanced-tool-use
So these numbers are at least 7 months out of date. Why is this being posted now?
Besides people with positions relevant to the field I'm weirded out by most of the replies, isn't MCP effectively just a communication standard? Like the only difference between an MCP server and my Express webserver is the supposed logic on how it needs to communicate with the AI, why are we making such a big deal out of it? Eventually we'll all converge into some form of standard to link things to our LLMs and it's probably going to be based in some form on MCP, but I genuinely don't get what the big deal is
I use all three (MCP/CLI/API) based on what Claude excels at:
* CLI: GitHub & AWS it already knows how to operate the CLIs well. Even learned about a few new CLIs like 1Password's op which it volunteered one day.
* MCP: Supabase, Shopify etc. where the CLI would be non-obvious and the affordances from the tools/descriptions helps Claude maneuver.
* API: Sometimes it just knows an API exists and is able to call it directly with python/curl. I discovered from Claude the Pokemon ecosystem has a free API out there for example.
> Update: Since these measurements were taken, Claude Code has rolled out Tool Search with Deferred Loading, which loads MCP tool schemas on-demand and reduces context usage by 85%+. The context bloat described in Problem 1 is largely addressed for users on current Claude Code versions. The performance, debugging, and architectural arguments below still apply.
My mental model for MCPs is that it's like a Swagger/OpenAPI spec for LLMs. Point 2 doesn't make much sense in that context as it's describing MCP as a Swagger endpoint that's unstable.
Chrome/Ghidra MCP does have a tendency of crashing, but I'm not sure why this is. Is my way of thinking of MCP incorrect? If it really is a descriptor of how to talk to another tool, then why do they seem fragile at times? I feel like there's a gap in my knowledge somewhere.
MCP is still great if you're running AI in an environment that precludes a shell while needing dynamic tool discovery, but that's a narrow set. People are learning how useful it is to give AI access to a shell. If you're giving them a shell, may as well give them a CLI.
However, I don't think that's what is really hurting MCP, because it could evolve. What really killed it was the standards process and enterprise groups getting ahold of it. It went into spec writing and got adjudicated into uselessness all while enterprise authentication groups were figuring out the best angle to make money on it. I listened to a pitch from Okta on MCP and they wanted to charge out the nose for it for no good reason.
I've thought that skills and small scripts > MCP for quite a while now, tried out MCP in the early days (official ones, ones i made for scripts i already had), but they always end up using more tool calls/tokens than if i had just written a script + skill for claude.
personally if i have the need to move a skill/script, etc. to another one of my machines, i'll make a git repo for them (if they aren't already on git)
This was one of the first ideas me and my team had for sharing skills and scripts. The problem is this is a very "why Dropbox and not FTP" answer.
The second you utter the word git, you may have lost 90% of your audience - depending on their background, of course. MCPs are a lot more non-tech friendly
yeah it 100% depends on who you'll be sharing them with, for me its just myself and a couple agents i have on a dedicated machine so git is ideal to keep versions matching when i update something on my daily driver
I think those are solvable problems. E.g. wrap mcp in skill or seperate forked (non context eating) call to smaller model to ask which mcps are applicable. Iet probably does this. Honestly I have not had issues with MCPs where I felt compelled to debug them.
MCPs are very useful when you don't have a CLI or you do but the MCP can handle auth like a proxy to something (e.g. Splunk). Or just for the USB-C analogy she gave.
I was writing MCP servers, now I just write tools for agents to consume. It's often easier to simply write the tool you need and suggest to it to look at the tool to do that thing.
I was also surprised to find out Claude knew how to use the gitlab api with pointing it at the token var in the environment. But for corporations it might make more sense to use a cli to keep the secrets separate from the agent.
> Using existing CLI directly: No context wasted on tool definitions
Can someone explain this to me? I've seen claude code try to run a not-well-known package and it basically shot in the dark a command, noticed that failed, then ran the help command for the cli tool to get a list of commands and what they do.
How is that different than passing the tools with an MCP? Like how are we saving context?
The usual problem is companies write an MCP server with 50 different tools, and each one has a schema, description, etc. Say each tool is 150 tokens, that's 150 * 50, or 7500 tokens, dumped into the beginning of every session. Compared to a text file that gets loaded on demand with command-line tool examples, so you still get close to the same amount of context, but you can control what tool definitions you pull in.
The other thing is the agent gets the entire MCP API response dumped into context as a tool response in JSON, which can be a lot. Compare that to shell commands where agents often `head` or `tail` or `grep` the response (which I kinda hate, but it does save tokens).
It also depends on whether the agent loads them on-demand or not (most modern agents do), and whether your MCP has a ton of tools or not. If your MCP only has 2 tools, and the responses aren't big, it's really not that much context.
The other thing that doesn't get talked about is the non-determinism of shell one-liners. There is a lot more non-determinism in shell tool calls; the AI can mess up commands, options, arguments. It can incorrectly filter output, miss output, miss return status, which results in re-running calls, polluting context, making results worse. Compare that to MCP calls which are more likely to succeed because they have a schema, well-defined errors, etc. Do you want less token use or more reliable results?
The thing is, you don't have to pick a side. I personally use both MCPs and CLIs at different times in different ways. Often I'll have the AI write a small script to do many calls (sometimes with tools, sometimes with libraries) which saves tokens, allows me to review, and is more deterministic.
So what's this saying? Rather than trust the llm to query external tools via mcp you should handle the external queries yourself? Otherwise the llm wastes a bunch of queries?
Man I wish I could downvote stories. There needs to be some way to push back against dark patterns in writing, like clickbait.
Clearly MCP is not dead, as the article itself says. But the article lies in order to play on human sentiment/heuristics and steal your attention. It's like shouting fire in order to get people to run over to see your business.
There's a fix for the context which involved an mcp search and execute gateway. Essentially the mcp server queries for desired capabilities, gets search results with execution and requirement details and fires off the actual mcp as subcalls:
You can call it "rag for mcp". I was pushing it hard a few months ago and nobody seemed to care but I'm all in if the timing has caught up to the tech.
It's nontrivial effort: basically a giant survey of all the mcp servers, running inference over them to figure out how to instrument them, cross referencing to make sure they are the "official" sources (or at least the ones that search engines think are) then using qdrant to do embeddings and reranking and offering it for free.
If people have become interested I'm all in. I'll bring the infra back up. I just don't want to spin my wheels on dead end streets.
The value proposition is solid, the problem is real, this fix works, it's fast, it's free, and people give exactly zero shits. I dunno...
While the title is quite obnoxious, the author is right.
I don't think that anyone would argue against standardizing training for any model on ways of invoking tools through specific output templates (with MCP being an extension of that). However, the question is what is the best way of having the model use those tools? There are 2 options
1 - Encode actual functionality during training, let the model figure out how to use available tools to do what it needs to. Latest Claude models are a good example of this, when editing files if it encounters issues with the under the hood tool, it will write a bash python command to edit the file
2 - Describe functionality in instruction context. This allows you to define complex sequences of things to do, but at the risk of the model losing context as the conversation continues.
3 - Use tool calling, where every request gets an available tools section appended to it, and define the complex functionality in the static code (whether its local tools or MCP servers)
Ideally, if we are pushing towards smarter models, the answer is between 1 and 2, where you have a model that only has access to be able to run shell commands, and some memory that it can reference on sequences of shell commands to run. An MCP invocation is then a simple echo jsonrpc pipe to local executable or a curl command. Eventually, its probably worthwhile to have your LLM run in a CPU like sandbox where it can execute arbitrary assembly commands from sequences stored in memory to do what it needs to do.
Until then, 2 and 3 are really what we have for adapting with current frameworks.
The thing that all these "MCP is dead" posts are missing is that whether or not MCP is used as a transport protocol is actually completely irrelevant.
The reason MCP isn't dead is because practically ~every company on the planet is building an MCP server. I know this because we interact with all of them. Most of these companies don't have a CLI. Many of these companies don't even have an external API! And yet, they're all building MCP servers.
And that's why MCP is not only not dead, but more important than ever.
Maybe we will turn every MCP server into a CLI under the hood. Maybe we'll use code mode. Maybe we'll implement tool search.
All of those are just implementation details to the much more important point: our AI agents are getting access to services they otherwise would never have had access to.[0] That's what matters.
So, is MCP dead as a direct communication layer for models to speak to? Maybe, maybe not. Is MCP dead as a protocol? Hell no, couldn't be further from the truth.
[0]: Although I will say the Codex app's computer & browser use features have made this statement a lot weaker than it used to be. If you haven't tried them yet—they're mindblowing.
If all of these companies spent equivalent time writing a CLI for agents to consume as they spend on MCP servers, would they be any worse off in terms of agents being able to interact with their products?
They have different tradeoffs.
That’s covered in the article: a human can modify the commands generated by the agent, or vice versa, to debug problems or transfer knowledge.
AI is a bandwagon tech, a lot of people will 'build because others are' adhering to an ostensible standard.
Most of the people that I know are moving away from MCP in favour of skills where the advantage of MCP goes away if the REST API is clear enough.
Also - I'm sorry to say but MCP management on Codex (and Claude) is just really bad. Everything from discovery, to management, to context window, to documentation - it feels unfinished as a 'feature' even if the protocol is supposed to be narrow.
1) I have a big popup and yellow warning every time a window is opened or a sub agent is launched warning me that 'SkySomething Computer Use' does not work. I had to Google to find out that has something to do with Codex MCP. So already the externalizations of problems, resolutions ... not very well done.
I'm not even sure what to do - and I'm honestly not interested in 'fixing' something I didn't cause, I'm not sure of, and don't want to deal with.
2) Just listing the current MCPs, knowing really what the are for (clearly, concisely) is hard.
This is what you get if you type /mcp in Codex:
/mcp
What's that supposed to mean? What is 'codex_apps'?As presented - it resolves to 'nonsense gibberish'. Those are things that I did not even install.
Do you expect people magically know what 'codex_apps' is?
Here is what 'AGI!' Codex 5.5 answered when I asked about 'codex_apps' is:
====
" codex_apps appears to be Codex’s own internal cache/tooling area, not part of J1 (my project).
"I only found it under .codex, e.g.:"
" I did not find it referenced by the J1 source. So unless you saw it somewhere specific, treat it as Codex runtime metadata for app/tool integration, not project code."
====
So even Codex itself has no idea what it's own MCP tools are, and after a full '1 minute of thinking' on 'xhigh' it responded with nonsense.
This whole experience fundamentally deflates my perception of AU, OpenAI, Codex and MCP.
This is supposed to be the 'future' but it feels like 1982 dialup.
This is where 'traditional UX' really starts to show it's value obviously, but you really need to consider enhancing this experience, possibly with some traditional ux mechanisms.
3) Knowing the 'state' of the MCP is totally opaque. Is the 'MCP server' running? Can I restart it? That might be outside the scope of 'Codex' but you're offering the product so all of the underlying stuff is essentially 'your responsibility' as well at least from a UX perspective. Why isn't the 'state' of the MCP listed.
4) How can I not just easily enable/disable individual MCPs so they don't chew up context?
5) How can I not discover MCPs from codex itself, so that I can find solutions to problems? MCPs are all a bit different, and awkward to install and manage. Like with VS Code, we can 'discover plugins'. Even from the Web we can search and discover plugins.
While I realize that most of this rant is oriented around MCP tooling management, and not the standard, I do feel that these issues are 'fundamentally at the crux' of the situation.
Our team has moved away from MCP into Skills - and after doing so, it's hard to see why MCP is going to be valuable - other than plausibly as defining some 'jon calling conventions'.
There's a lot of obvious things to improve, please do that.
That the 'smartest people in the world' have $100 Billion and are are totally scattered on so many issues completely blows my mind but it speaks to how systems are organized.
They don't need to hire anyone for this stuff, they need to have some basic product discipline and prioritize it, that's all.
If they don't do that, all the money in the world and all the best product people wouldn't be able to help.
I totally respect that 'Codex is young' - but it's been kind of a year now. That's a long time - and AI is supposed to 'accelerate time scales' and make people 'super productive' remember?
I hope they improve.
MCP is essentially just JSON RPC with a few special fields that must be included. I have reservations about JSON RPC, but there needs to be some 'service discovery' layer for LLMs to interface with.
It needs to be available in places like websites, desktop applications, backend services, etc. The CLI is only one place that these systems interface with.
Whatever you replace MCP with will be in a similar shape even if you specify a different communication protocol or different fields for tool discovery.
'Skills' should all be based on MCP, they should load on demand, be very easily manageable and discoverable by humans and by AI, and then it would work
The scope was too narrow, given how it ended up being applied.
If they layer something on top of it, it may yet be revived.
Essentially it's anything that _could_ be on a dashboard, but _might_ be accessed conversationally via an agent.
So these numbers are at least 7 months out of date. Why is this being posted now?
* CLI: GitHub & AWS it already knows how to operate the CLIs well. Even learned about a few new CLIs like 1Password's op which it volunteered one day.
* MCP: Supabase, Shopify etc. where the CLI would be non-obvious and the affordances from the tools/descriptions helps Claude maneuver.
* API: Sometimes it just knows an API exists and is able to call it directly with python/curl. I discovered from Claude the Pokemon ecosystem has a free API out there for example.
i personally was anti-MCP but they just work better in terms of tool search than a CLI, especially with the idea of tool nudging
pretty much
Chrome/Ghidra MCP does have a tendency of crashing, but I'm not sure why this is. Is my way of thinking of MCP incorrect? If it really is a descriptor of how to talk to another tool, then why do they seem fragile at times? I feel like there's a gap in my knowledge somewhere.
However, I don't think that's what is really hurting MCP, because it could evolve. What really killed it was the standards process and enterprise groups getting ahold of it. It went into spec writing and got adjudicated into uselessness all while enterprise authentication groups were figuring out the best angle to make money on it. I listened to a pitch from Okta on MCP and they wanted to charge out the nose for it for no good reason.
The second you utter the word git, you may have lost 90% of your audience - depending on their background, of course. MCPs are a lot more non-tech friendly
Its not that hard.
MCPs are very useful when you don't have a CLI or you do but the MCP can handle auth like a proxy to something (e.g. Splunk). Or just for the USB-C analogy she gave.
I was also surprised to find out Claude knew how to use the gitlab api with pointing it at the token var in the environment. But for corporations it might make more sense to use a cli to keep the secrets separate from the agent.
Can someone explain this to me? I've seen claude code try to run a not-well-known package and it basically shot in the dark a command, noticed that failed, then ran the help command for the cli tool to get a list of commands and what they do.
How is that different than passing the tools with an MCP? Like how are we saving context?
The other thing is the agent gets the entire MCP API response dumped into context as a tool response in JSON, which can be a lot. Compare that to shell commands where agents often `head` or `tail` or `grep` the response (which I kinda hate, but it does save tokens).
It also depends on whether the agent loads them on-demand or not (most modern agents do), and whether your MCP has a ton of tools or not. If your MCP only has 2 tools, and the responses aren't big, it's really not that much context.
The other thing that doesn't get talked about is the non-determinism of shell one-liners. There is a lot more non-determinism in shell tool calls; the AI can mess up commands, options, arguments. It can incorrectly filter output, miss output, miss return status, which results in re-running calls, polluting context, making results worse. Compare that to MCP calls which are more likely to succeed because they have a schema, well-defined errors, etc. Do you want less token use or more reliable results?
The thing is, you don't have to pick a side. I personally use both MCPs and CLIs at different times in different ways. Often I'll have the AI write a small script to do many calls (sometimes with tools, sometimes with libraries) which saves tokens, allows me to review, and is more deterministic.
Clearly MCP is not dead, as the article itself says. But the article lies in order to play on human sentiment/heuristics and steal your attention. It's like shouting fire in order to get people to run over to see your business.
> Problem 1: It Devours the Context Window
Don't harnesses support progressive discovery these days?
Claude (200K).... GPT-4o..........?
> every MCP server adds a process layer between the LLM and the underlying API
But a CLI doesn't?
------------------
> Measurement: Tool Definition Sizes
> MCP Server: Linear, Notion, Slack, Postgres
Oh, so these are the MCP servers that are examples of context bloat we're going to replace! Later in the article:
> At Quandri we use all three approaches side by side...
> MCP for services without a strong CLI (Slack, Linear, Notion)
https://github.com/day50-dev/mcp-search-and-run
You can call it "rag for mcp". I was pushing it hard a few months ago and nobody seemed to care but I'm all in if the timing has caught up to the tech.
It's nontrivial effort: basically a giant survey of all the mcp servers, running inference over them to figure out how to instrument them, cross referencing to make sure they are the "official" sources (or at least the ones that search engines think are) then using qdrant to do embeddings and reranking and offering it for free.
If people have become interested I'm all in. I'll bring the infra back up. I just don't want to spin my wheels on dead end streets.
The value proposition is solid, the problem is real, this fix works, it's fast, it's free, and people give exactly zero shits. I dunno...
One day I'll figure it out, hopefully...
scrolls down the page...
> So is MCP really dead? Not entirely
sigh...
While the title is quite obnoxious, the author is right.
I don't think that anyone would argue against standardizing training for any model on ways of invoking tools through specific output templates (with MCP being an extension of that). However, the question is what is the best way of having the model use those tools? There are 2 options
1 - Encode actual functionality during training, let the model figure out how to use available tools to do what it needs to. Latest Claude models are a good example of this, when editing files if it encounters issues with the under the hood tool, it will write a bash python command to edit the file
2 - Describe functionality in instruction context. This allows you to define complex sequences of things to do, but at the risk of the model losing context as the conversation continues.
3 - Use tool calling, where every request gets an available tools section appended to it, and define the complex functionality in the static code (whether its local tools or MCP servers)
Ideally, if we are pushing towards smarter models, the answer is between 1 and 2, where you have a model that only has access to be able to run shell commands, and some memory that it can reference on sequences of shell commands to run. An MCP invocation is then a simple echo jsonrpc pipe to local executable or a curl command. Eventually, its probably worthwhile to have your LLM run in a CPU like sandbox where it can execute arbitrary assembly commands from sequences stored in memory to do what it needs to do.
Until then, 2 and 3 are really what we have for adapting with current frameworks.