As you all probably know by now, I kind

of love bullying Anthropic. As great as

their models can be, they have their

weird behaviors and things that I just

don't love about how the company

operates. One of those things in

particular is the way they treat open

source. I've never been super fond of

it. But they recently acquired Bun, the

JavaScript runtime, that I do dearly

love, both the framework and the runtime

itself, as well as the team building it.

And it seems like with this acquisition,

a new focus on open source is happening,

which is why this announcement, as crazy

as it is, is actually quite cool. And

Thorback just donated the model context

protocol to the new Agentic AI

Foundation. If you're curious what this

is, spoiler, it's a new fund that was

just formed under the Linux Foundation.

Yes, really. The Linux Foundation are

now the owners of the model context

protocol. This is a fascinating new

development. both the creation of this

new agentic group within Linux

Foundation, but also MCP becoming a more

open standard that exists in an external

entity instead of just being part of

Anthropic. This is super super exciting

stuff. Projects like this are only

possible because wellunded companies

like Anthropic Block, OpenAI, Google,

Microsoft, and Amazon support them. But

we need support too. So let's do a quick

word from today's sponsor. Today's

sponsor is different and I want to give

you guys a bit of history on them. They

started their journey as the best

possible API for giving context on your

codebase to different AI agents. So if

you're trying to build your own AI that

deeply understands your codebase, their

product built around like modern grap.

It turns out that's useful for a lot of

other things though. In particular, code

review. If an AI agent knows everything

about your codebase, it can actually

give you really good feedback when

you're trying to change it. The sponsor

is Grapile and I want to show you what

this looks like in practice. This is a

code review on a real repo. It's Physics

Nemo by Nvidia. like actual real big

legit project. Grab Tile left a review

here. Gave it a low confidence score, a

three out of five. But god, I love this

feature so much. The way that it will

tell you how it feels overall, it's just

so clear and easy to know like, do I

need to give this another look or not?

Confidence score is a three, but the

reason why is where it gets interesting.

Gravile called out that there is a

precision calculation issue on this

particular line. The piece that's

interesting here is it can only know

this with that codebase context. that

the way that these values are calculated

other places is to put both of these

values underneath when it's doing this

division. One of the maintainers claimed

otherwise saying that the true positives

over predicted positives can be greater

than one which is correct in their

understanding. GPile responded which is

what makes it so different. Looking at

your concern, I need to clarify the

issue. The formula on line 200 is indeed

wrong but true positives over predicted

positives cannot be greater than one

when calculated correctly. The problems

on line 196. It's calculating the count

of all true class zero labels which

should be called actual positives not

true positives or precision true

positives is the same. You get the idea.

It's a really good fellow code reviewer

and as we could see here Kelvin

confirmed as much. Turns out I was wrong

and overlooked the true positives as the

intersection. I fixed this and also

flipped the positive as label one. There

you go. If all of your code reviews are

still being done by humans, you're

wasting all of those humans time. Pull

in Gravile today at soyv.link/graphile.

donating the model context protocol and

establishing the Agentic AI Foundation.

Today we're donating the MCP to the

Agentic AI Foundation, AIF, a directed

fund under the Linux Foundation, which

is co-founded by Anthropic, Block, and

OpenAI with support from Google,

Microsoft AWS Cloudflare and

Bloomberg. One of these things is not

like the others. Bloomberg's code

discipline in particular is actually

really good. Their TypeScript stuff is

incredible. I thought MCP was a little

more than a year old. Is that really

right? No, 24. Wow. Yeah, it's one year

later. That's crazy. Almost to the dot.

One year ago, Anthropic introduced MCP

is a universal open standard for

connecting AI applications to external

systems. Notably, connecting and

external systems, implying that the

server that you're connecting to is

expected to be live and maintain a

connection during the entire session,

which makes it really annoying to deploy

and adopt. So, a couple cool new

companies working on figuring out

solutions to that now. But yeah, the the

way MCP is implemented is really focused

on making the agent side as easy as

possible, not on making the developer

implementation side on the other end as

easy as possible or really viable at

all. MCP's achieved incredible adoption.

There are over 10,000 active public MCP

servers covering everything from

developer tools to Fortune 500

deployments. Across platforms, MCP has

been adopted by ChatGBT, Cursor, Gemini,

Copilot, Visual Studio Code, and other

popular AI products. Enterprise grade

infra now exists with deployment support

for MCP from providers including AWS

Cloudflare, Google Cloud, and Azure. It

does make sense that the big clouds

would want to be part of this. It's also

nice having them supporting Linux

Foundation more. You're not already

familiar with Linux Foundation. No,

they're not like the people who make

Linux. It's a little more complex than

that. The Linux Foundation is an attempt

to take a lot of like open- source work

and orchestrate it while also getting

money in memberships from big companies

to try and make sure the direction of

all of these important big open source

projects continues to make sense for how

everybody is using them. They have a

fascinating set of projects including of

course the Linux kernel, risk 5 and

pietorch. So the things underneath them

are really important and fascinating.

You guys remember the chaos around

Reddus? Remember when they changed the

license and suddenly all the other

providers that hosted it had to like not

do that or use an old version? When that

happened, both AWS and Microsoft put a

lot of money in in order to fund the

development of a fork called Valky that

included a bunch of the original

maintainers of Reddus as well as people

maintaining it at Azure and at AWS. They

then instead of trying to run this

independently and separately, donated it

to the Linux Foundation alongside I'm

sure more donations. The result being

the Linux Foundation now owning a good

open standard for Reddit style use cases

that is truly open and will stay open. I

don't think there's any example of

something that went to the Linux

Foundation to be a well-maintained open

source project then locking down which

is a big reason why Anthropic would be

interested in this. There's always fear

that an open standard could get closed

and the way you counter that is going

really far alongside it which is what

happened here. Since there's now concern

about Reddus potentially changing its

license and the rugpull that can occur,

the alternative to that isn't another

thing that can be rugpulled, it's to

throw it at Linux Foundation so it will

be there hypothetically speaking

forever. A lot of the things Linux

Foundation maintains are things that

just have existed forever and probably

will as well. It's also worth noting

that the Linux foundation has other sub

foundations that they help operate as

well like the cloudnative computing

foundation, openjs foundation, the open

source security foundation and more.

There's a lot of these foundations that

they help manage and run is open search

their attempt to do uh elastic search

because I know that they had an elastic

search alternative at some point as

well. Regardless, a new one has just

been dropped which is the agentic AI

foundation which is one of the Linux

foundation projects. The AIF provides a

neutral open- source foundation to

ensure the critical capability evolves

transparently, collaboratively, and in

ways that advance the adoption of

leading open source AI projects. It is

kind of funny that the Open Linux

Foundation Agentic AI foundation is

funded by the companies doing all of the

closedweight models. Like, has Anthropic

released any openweight anything in the

last like 3 to four years? I'm pretty

sure no. Opening just recently did one,

but you would expect a Meestrol or

DeepSeek to be here, but they also don't

have the funding to justify it. It is

cool to see though. They now have MCP as

well as Goose, which is an open source

accessible AI agent that goes beyond

code suggestions. Interesting. And

Agents MD they also own. Very

interesting actually. I did not realize

that Agents MD was big enough to go to a

foundation like this, but that does make

sense. And I've seen more and more cool

uses of Agents MD files. So, this is

cool. Oh, interesting. Goose is by

Block. If you're not familiar with

Block, they are a company that owns a

lot of different things. They own

Square, Cash App. They bought Title,

which uh is close to my heart, as well

as a few other things. It was founded by

the original founder of Twitter. Coolish

company. No issues with them. It's cool

seeing them doing more open source stuff

like this. Wait, no way. Yeah, I know

Roselle. She's great. That's so cool. I

had no idea she was at Block now. That

makes a ton of sense. Good for her. Good

stuff. Local AI agent autonomously

engineering tasks seamlessly. This is a

good set of things. It's cool to see

something like this forming. Now, it's a

weird set of projects. You have a

markdown file standard. You have an

actual big open-source AI agent toolkit

and then MCP which is a protocol for

interfacing with agentic stuff. Very

interesting set of things especially

when you look at the set of members that

they have. the block membership makes

much more sense now. Seems like they

kind of want to pawn goose off there,

but also the block founders have always

been very very pro- opensource things.

So, it does make a lot of sense there.

Back to the article.

Oh, they have a whole fancy little

timeline of the history of MCP here.

Early adoption from cursor of the AI

tools back in December to February. MCP

was presented at the AI engineering

summit. There was 2,000 public MCP

servers at the time. A second spec came

out which included enterprise off HTTP

streaming multimodal content but also

the enterprise o solution they picked

was lackluster at best. It's my

understanding that they used an existing

HTTP standard for O that had never

really been implemented by anyone. That

kind of sucks and as such it was not

something many people did. It's also

funny that a thousand people register

for the MCP night in San Francisco

because that's happening again tonight

and I will be there. That's why I'm

ending stream early. At this point, they

had 4,000 public MCP servers. Then June

to July, the third spec came out with

structured outputs, which is essential

to have the MCP actually output things

into consistent shape. Interactive

prompts as well as hardened security.

They also introduced the MCP directory

and 2,000 people took part in the agents

hackathon. Then enterprise MCP hosting

started happening on Google Cloud, Azure

and Cloudflare. MCP DevSummit sold out

in London. This is also around when we

started to see these like the push back

on the idea that MCP is a thing that is

just how models will do application

development and application style work

make less and less sense. And that's

when I did the videos about the

Cloudflare article about turning MCPs

back into code and then letting the

agents write code to execute instead.

Watch those videos if you haven't and

you want to better understand the things

I don't like about MCP as well as how we

can take advantage of the standard more

effectively. This is interesting though.

I also don't think the growth is that

impressive from like the first two

months they got 2,000 servers and then

over the next 10 months they got another

8,000. That's relatively linear growth.

I'm once again going to run a poll. Do

you use MCP? Yes, a lot. Greater than

five servers? Yes, a bit less than five

servers. Not really. Nope. I'll change

it to tried but churned. I should have

put a just contact 7 option. I'll do

another poll after for that. Yeah, gave

up on Context 7, which is the only MCP I

was trying to use after the last poll in

chat. According to my audience, 56% of

people don't use MCP, 16% tried and

churned, 25% use it a bit, and 4% use it

a lot. It's kind of crazy how close the

a bit of use and tried and churned are,

but this is my experience, too. I just

have not found many use cases for MCP in

my day-to-day usage. Yeah, the use case

I see for it, funny enough, is much

stronger locally. I've yet to find many

MCPs that are useful if you run them

remotely. Like there's the Mac MCP that

lets you control your Mac. Of course,

Super Memory are the ones doing it. This

is the Mac MCP is pretty cool. The point

of it is that you can control features

on your Mac via MCP. There's a bunch of

TypeScript code written to do things on

your Mac exposed over a local server

that your AI agents can call and do

things like send texts or update your

notes, touch your contacts. And here you

can see in Daria's example, he tells it

to send a message, this is a test to

somebody. And now it's running behind

the scenes on his Mac doing the things

that he requested, which is pretty cool.

But those are the use cases that I think

are the most valid right now due to the

stranges of how the protocol is

implemented. There's also like the

Ableton MCP. If you're a music nerd,

it's pretty cool. Ableton MCP is another

very fun one. It's for controlling the

Ableton Live digital audio workstation.

It's a very common tool for producing

music. And Ableton MCP lets you modify

your project in Ableton using bottle

context protocol. But once again, since

you have to run a full dedicated server

for this to work, you have to set up the

smiththery CLI, install this MCP, and

specify the client is claude. You

probably have to install UV as well

because it's all a Python server and

then spin it up and run it locally in

the background because like MCP can't

execute things that it doesn't have. It

is effectively an HTTP server that is

exposing specific behaviors to a model

and allowing for it to stay connected to

this server so it can do things and get

data and make changes. Another notable

thing about these examples is none of

them require complex off stories because

O is one of the most annoying nightmares

across all of these things. I don't even

know if you can off an MCP server and

cursor. Oh, fun. You have to hardcode

the tokens or you can pull them from

environment in the definition that you

paste in for the project. Fascinating.

Yeah, it's cool. They do support it now.

You also have to manually toggle the

features that you do and don't want it

to have access to on and off. Fun. You

couldn't tell I don't use MCP and cursor

a whole bunch because I don't use MCP a

whole bunch. You can also return images

as long as they're base 64. Fun, man.

What an interesting standard. Back to

the article because there more important

pieces here. We're continuing to invest

in MCP's growth. Clouds are directed

with over 75 connectors which are

powered by MCP and they recently

launched the tool search and

programmatic tool calling capabilities

in their API. I like that they worded

this carefully so they don't have to

call it the tool search tool as they

have been constantly. They did tool

search and programmatic tool calling

capabilities which helps with production

scale MCP deployments handling thousands

of tools efficiently and reducing

latency and complex agent workflows.

Yeah, that's a big step. It feels like

Anthropic recognizes that improving the

standard is not necessarily within their

area of expertise and instead of trying

to patch it again and again, it is much

more efficient for them to hand it off

to a better open standards committee

that can be designed. I know design by

committee is terrible, but for things

that are standards that are going to be

adopted by many, it makes a lot more

sense. make the standard external so it

can be iterated on, developed and owned

safely, openly and externally and

continue building the new features on

top. It definitely seems like their

recent interest has less been in

changing how MCP works and more in

building things around it. Like if we go

back to the timeline, they got early

adoption, they presented it, they

released the second spec which included

the O stuff. Then in June and July, they

did the third spec with structured

outputs, interactive prompts, and harden

security. And since then, a lot of the

new stuff isn't really part of the spec.

Like discovery isn't really part of the

spec for MCP. Async tasks are, but are

more like an implementation detail on

the thing consuming. And agentic

sampling, I don't even know what they

mean by that. Interesting. Sampling

allows for the server to reverse request

back to the model to do a thing. So if

you're using Opus 4.5 and it uses the

MCP for I don't know doing a a query

against your database to get some data

and then your code realizes because you

wrote the code this way that it has to

have the right SQL to generate it can

forward back that request to the model

to generate the SQL string that it has

to run. I've never seen anyone using

this. This is really interesting too.

There's a hints section which is an

array of arbitrary JSON objects by the

looks of it, but they're separately this

intelligence priority and speed priority

things that are top level keys which

means that they're probably part of the

spec,

man.

Yeah, you can see that the chaos that is

this diagram. The server initializes the

sampling with the sampling create

message. A human can review that

request. They're presented the request

for approval. they could request a a

modify or just approve it and then it

forwards that request to the LLM,

returns a generation. You can review the

response. The response is presented to

the human for approval and if they

approve it, it gets sent back to the

client which then sends it back to the

server to continue. Man,

when I say it feels like all of this is

chaotically overengineered,

do you understand?

It is also kind of chaotic that the the

client in this case is something like if

you use the cloudi desktop app or if you

use cursor the server is sending a

request to the client that the user then

approves so the client can then resend

it to the LLM. This weird graph of where

everything lands is just strange. It can

also request images back and audio.

Does

this feels like a standard looking for a

problem still? Like who's going to

implement this? What tools exist right

now where if I wrote my MCP server to

then request an image back to the client

that will handle that properly? I just

don't think there's going to be many.

Oh, and then yeah, the these are

standard cost priority, speed priority,

and intelligence priority. Cost

priority. Higher values prefer cheaper

models. Speed higher prefers faster. And

intelligence, higher prefers more

capable. And the values, by the way, are

between zero and one. Great. Then the

hints allow servers to suggest specific

models or model families. Again, they

had cloud 3 sonnet and then claude as a

fallback.

Yeah, interesting. So, yeah, that is a

new part of the standard. is not one

I've heard anyone talking about, but

they did add something. But it does

generally feel like the standard in the

spec has less been Anthropic's focus and

more the things they're building around

it, which makes the punting of it off as

they build new tools on top of it make

more and more sense. It is also like to

be fair a really good faith play to say

we're not just building MCP for our own

use cases and tricking you all to build

on top of it. They do clearly want MCP

to be a winning open standard. I have no

reason to believe that any of their way

of doing this is malicious. I still have

my skepticism around MCP, but I don't

think they're doing this because they're

dumb or evil. I think they do genuinely

believe this is an important thing to

exist and that it is the best chance we

have at an open standard for AI to do

things on real systems. Still kind of

crazy that they're making this so open

and cloud code is still closed source

and they've sent as many DMCAs as they

have to GitHub about people who shared

the source maps for cloud code when they

leaked them in their own package when

they deployed it. Just saying. The Linux

Foundation and the Agentic AI

Foundation. The Linux Foundation is a

nonprofit organization dedicated to

fostering the growth of sustainable open

source ecosystems through neutral

stewardship, community building, and

shared infrastructure. It has decades of

experience stewarding the most critical

and globally significant open source

projects including Linux, Kubernetes,

Node.js and PyTorch. Yeah, Node is

technically part of it too which is

nuts. Importantly, the Linux Foundation

has a proven track record in

facilitating open collaboration and

maintaining vendor neutrality. Yes,

absolutely yes. Node runs just as well

everywhere, which is kind of crazy if

you think about it. Like you all know

how much people complain about Nex.js

being easiest on Versel. Node works

everywhere other than Cloudflare kind

of, but you get the point. Node is a

very much supported standard that works

equally well pretty much everywhere.

It's really cool that that's proven.

Same with Linux kernel. I would imagine

a significant portion, if not the

majority of the money that Microsoft

makes on Azure right now is going

through Linux systems. Agentic AI

Foundation is a directed fund under the

Linux Foundation co-founded by Anthropic

Block and OpenAI with support from all

the other companies I mentioned before.

The AIF aims to ensure agentic AI

evolves transparently, collaboratively,

and in the public interest through

strategic investment, community

building, and shared development of open

standards. The idea of agentic AI

evolving transparently by making MCP an

open standard is an interesting thing. I

would imagine that transparent evolution

of AI and agentic work would involve

more transparency around how the models

are being trained, which none of the

labs are willing to do. Some deeply

skeptical part of me wants to see this

as an attempt to look more transparent

when they are more than ever closing off

their research and work so people can't

copy their details. But I do really want

to read this in good faith cuz there's

no reason not to at this point at least.

I'll be keeping my eye on it. And now

the donating MCP section. Anthropic

dating the model context protocol to the

Linux Foundation's new agentic AI

foundation where it will join Goose by

block and agents MD by OpenAI as

founding projects. I didn't know agents

was open AI. That does make sense.

Bringing these and future projects under

the AIF will foster innovation across

the agentic AI ecosystem as well as

ensuring that these foundational

technologies remain neutral, open, and

communitydriven. The MCP's governance

model will remain unchanged. The

project's maintainers will continue to

prioritize community input and

transparent decision-making. There is a

lot of work on this. I've seen so many

issues that are hundreds upon hundreds

of replies deep as well as their shared

Discord server.

They are trying to make it better. I've

had a lot of people, like more so than

almost anything I've complained about,

I've had a lot of people who are part of

the MCP spec and standards committee

replying and reaching out trying to

figure things out, especially the off

things. It seems like the off stuff's a

total disaster. The future of MCP open

source software is essential for

building a secure and innovative

ecosystem for agentic AI. Today's

donation to the Linux Foundation

demonstrates our commitment to ensuring

MCP remains a neutral open standard.

We're excited to continue contributing

to MCP and other agentic AI projects

through the AIF. Yeah, this is cool. If

they want to prove that it's an open

standard, they did it right. But it also

kind of seems like Infopa can only have

one open source project at a time. And

now that they have Bun as their big open

source project, MCP has been moved away.

Openai's article about this foundation

has even more interesting stuff. I'm

going to read this a little too. Similar

blurb here. It is cool that they're

calling out that they're working with

Anthropic and Linux Foundation block all

of them. I've never seen OpenAI call out

Anthropic as a supporter and Anthropic

call out OpenAI in matching articles the

same day. Weird to see this much

alignment and genuine collaboration on

the standards here. Why open standards

matter? Developers are rapidly adopting

AI to build more capable agentic systems

from coding assistance to workflow

automation and customer service agents.

In 2025, these systems have begun to

move from prototypes into tools that

handle real work in business and

consumer settings. We believe the

transition from experimental agents to

real world systems will best work at

scale if there are open standards that

help make them interoperable, like a

standard for a markdown file. Open

standards make agents safer, easier to

build, and more portable across tools

and platforms. And they help prevent the

ecosystem from fragmenting as the new

category matures. As more agents begin

handling real responsibility, the cost

of fragmentation increases. Without

common conventions and neutral

governance, agent development risks

diverging into incompatible silos that

limit portability, safety, and progress.

You were around for the early web days

between like the 15 different attempts

to do dynamic web apps. I see why

jumping on this early is a thing they're

all focused on. Agents SDK isn't open

source, is it? What is the license? MIT.

Is it standard MIT? It is. Cool.

OpenAI's history with the agents SDK,

apps SDK, and Agentic commerce protocol

is cited here as proof that they really

want to contribute to these building

blocks and open standards and codec CLI

was a pretty baller move. When they

dropped that as an open- source project,

I was surprised. I did not see that

coming, especially after how strictly

cloud code has stayed closed source.

It's a baller move on their part. And

then they did the GPTO OSS models, which

believe it or not, I'm still using

GPT120 bill quite a bit. It runs really

well on my laptop, too. Is the apps SDK

open source as well? There's no way.

These are just examples. That's what I

thought. The apps SDK is not open

source. Also, to give Enthropic a little

more grace here, the apps SDK and the

idea of apps in chat GPT was so hyped

when they announced it and I haven't

seen anyone use it for anything like not

a single thing. Apparently, the UI

portion is open source, but yeah, no

one's using it. I don't care. They've

been early adopters and core

contributors to MCP incorporating it as

a foundation for connectors and apps in

Chat GPT. Just last week, they announced

a collab with Enthropic and MCP UI to

extend the apps SDK to all MCP

developers through MCP apps. Oh, great.

They're combining the two standards I

don't like.

And they donated agents MD.

Agents MD has been adopted by more than

60,000 open source projects and agent

frameworks including AMP, Codeex,

Cursor, Devon, Factory, Gemini CLI,

GitHub, Copilot, Jules, and VS Code

among others. Jules mentioned that's

Google's agentic coding tool that nobody

knows about. There's also a bunch of

other things they're not calling out

here. Like I know that uh code rabbit

and grapile both use agents MD very well

as well which is super helpful during

code review steps. It's cool that we

don't have the problem that we have with

a JavaScript project like uh you know

the classic config hell. Look at all

these dot files. The envitign Ignore and

VMRC oxlint prettier ignore prettier rc

components json nip json agentmd all

these things. It is pretty cool that the

agent MD is now just one file that works

with everything and we're not going to

have as bad of config hell. Like there

was a bit where you had to have the

cursor MD cursor rules agent MD winds

surfs equivalent and like 15 other

things. Agents MD has been standardized

enough that like everyone uses it. That

said, different tools benefit from

different types of instructions which it

varies. Still overall very good thing

especially if more and more of these

tools adjust to follow the agent MD

standard better. This seems cool. I

really like this type of collaboration

and alignment. I think it leads towards

better things happening. I really like

the idea of the Agentic AI Foundation

where if a thing becomes popular enough

that is meant to be open, the Linux

Foundation can now own it as part of the

Aif. This makes a ton of sense. I wish I

had more negative to say cuz it's fun,

but this is a good thing. This is the

best future for MCP. This is a really

good thing for all of these standards

and it makes the future of Agentic AI

tooling much, much safer and potentially

much better. Curious what y'all think

though. Am I way too excited about this

or is this just another weird MCP thing?