Anthropic Gave MCP to the Linux Foundation  As you all probably know by now, I kindof love bullying Anthropic. As great astheir models can be, they have theirweird behaviors and things that I justdon’t love about how the companyoperates. One of those things inparticular is the way they treat opensource. I’ve never been super fond ofit. But they recently acquired Bun, theJavaScript runtime that I do dearlylove, both the framework and the runtimeitself, 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 crazyas it is, is actually quite cool. AndThorback just donated the model contextprotocol to the new Agentic AIFoundation. If you’re curious what thisis, spoiler, it’s a new fund that wasjust formed under the Linux Foundation.Yes, really. The Linux Foundation arenow the owners of the model contextprotocol. This is a fascinating newdevelopment, both the creation of thisnew agentic group within LinuxFoundation, but also MCP becoming a moreopen standard that exists in an externalentity instead of just being part ofAnthropic. This is super, super exciting. Stuff. Projects like this are onlypossible because well-funded companieslike Anthropic, Block, OpenAI, Google,Microsoft, and Amazon support them. Butwe need support too. So let's do a quickword from today’s sponsor. Today’ssponsor is different, and I want to giveyou guys a bit of history on them. Theystarted their journey as the bestpossible API for giving context on yourcodebase to different AI agents. So ifyou’re trying to build your own AI thatdeeply understands your codebase, theirproduct is built around modern graph technology. It turns out that’s useful for a lot ofother things, though. In particular, codereview. If an AI agent knows everythingabout your codebase, it can actuallygive you really good feedback whenyou’re trying to change it. The sponsoris Grapile, and I want to show you whatthis looks like in practice. This is acode review on a real repository. It’s PhysicsNemo by Nvidia, like an actual real biglegit project. Grapile left a reviewhere. They gave it a low confidence score, athree out of five. But, God, I love thisfeature so much. The way that it willtell you how it feels overall, it’s justso clear and easy to know, "Do IDo we need to give this another look or not? The confidence score is a three, but thereason why is where it gets interesting.Gravile called out that there is aprecision calculation issue on thisparticular line. The interesting part is thatit can only know this with that codebase context.The way that these values are calculatedin other places is to put both of thesevalues underneath when it's doing thisdivision. One of the maintainers claimedotherwise, saying that the true positivesover predicted positives can be greaterthan one, which is correct in theirunderstanding. GPile responded, which iswhat makes it so different. Looking at your concern, I need to clarify theissue. The formula on line 200 is indeedwrong, but true positives over predictedpositives cannot be greater than onewhen calculated correctly. The problemsare on line 196. It's calculating the countof all true class zero labels, whichshould be called actual positives, nottrue positives. Precision truepositives are the same; you get the idea.It's a really good fellow code reviewer,and as we can see here, Kelvinconfirmed as much. Turns out I was wrong. And overlooked the true positives as theintersection. I fixed this and alsoflipped the positive as label one. Thereyou go. If all of your code reviews arestill being done by humans, you'rewasting all of those humans' time. Pullin Gravile today at soyv.link/graphile. Donating the model context protocol andestablishing the Agentic AI Foundation.Today we're donating the MCP to theAgentic AI Foundation, AIF, a directedfund under the Linux Foundation, whichis co-founded by Anthropic, Block, andOpenAI with support from Google,Microsoft, AWS, Cloudflare, andBloomberg. One of these things is notlike the others. Bloomberg's codediscipline in particular is actuallyreally good. Their TypeScript stuff isincredible. I thought MCP was a littlemore than a year old. Is that reallyright? No, 24. Wow. Yeah, it's one yearlater. That's crazy. Almost to the dot. One year ago, Anthropic introduced MCPas a universal open standard forconnecting AI applications to externalsystems. Notably, connecting andexternal systems, implying that theserver that you're connecting to isexpected to be live and maintain aConnection during the entire sessionmakes it really annoying to deployand adopt. A couple of cool newcompanies are working on figuring outsolutions to that now. The way MCP isimplemented is really focusedon making the agent side as easy aspossible, not on making the developerimplementation side on the other end aseasy as possible or really viable atall. MCP has achieved incredible adoption. There are over 10,000 active public MCPservers covering everything fromdeveloper tools to Fortune 500deployments. Across platforms, MCP hasbeen adopted by ChatGBT, Cursor, Gemini,Copilot, Visual Studio Code, and otherpopular AI products. Enterprise-gradeinfrastructure now exists with deployment supportfor MCP from providers including AWS,Cloudflare, Google Cloud, and Azure. It makes sense that the big cloudswould want to be part of this. It’s alsonice having them supporting the LinuxFoundation more. If you’re not alreadyfamiliar with the Linux Foundation, no,they’re not like the people who makeLinux. It’s a little more complex thanthat. The Linux Foundation is an attemptto organize and support a lot of open-source work. And orchestrate it while also gettingmoney in memberships from big companiesto try and make sure the direction ofall of these important big open sourceprojects continues to make sense for howeverybody is using them. They have afascinating set of projects including, ofcourse, the Linux kernel, RISC-V, andPyTorch. So the things underneath themare really important and fascinating. You guys remember the chaos aroundRedis? Remember when they changed thelicense and suddenly all the otherproviders that hosted it had to eitherstop using it or use an old version? When thathappened, both AWS and Microsoft put alot of money in to fund thedevelopment of a fork called Valky thatincluded a bunch of the originalmaintainers of Redis as well as peoplemaintaining it at Azure and at AWS. Theythen, instead of trying to run thisindependently and separately, donated itto the Linux Foundation alongside, I’msure, more donations. The result beingthe Linux Foundation now owning a goodopen standard for Reddit-style use casesthat is truly open and will stay open. Idon’t think there’s any example ofsomething that went to the LinuxThe Foundation aims to be a well-maintained opensource project, which is a big reason why Anthropic would be interested in this. There's always fearthat an open standard could get closed,and the way you counter that is by goingreally far alongside it, which is whathappened here. Since there's now concernabout Reddus potentially changing itslicense and the rug pull that can occur,the alternative isn't anotherproject that can be rug pulled; it's tosubmit it to the Linux Foundation, so it willbe there hypothetically speaking,forever. A lot of the projects the LinuxFoundation maintains are things thathave existed forever and probablywill as well. It's also worth notingthat the Linux Foundation has other subfoundations that they help operate, such asthe Cloud Native ComputingFoundation, OpenJS Foundation, the OpenSource Security Foundation, and more. There are many of these foundations thatthey help manage, and OpenSearch istheir attempt to create an alternative to Elastic Search,because I know they had an ElasticSearch alternative at some point aswell. Regardless, a new one has justbeen released, which is the Agentic AI. The foundation, which is one of the Linuxfoundation projects, provides aneutral open-source foundation toensure the critical capability evolvestransparently, collaboratively, and inways that advance the adoption ofleading open-source AI projects. It iskind of funny that the Open LinuxFoundation Agentic AI foundation isfunded by the companies doing all of theclosed-weight models. Has Anthropicreleased any open-weight anything in thelast three to four years? I'm prettysure no. Opening just recently did one,but you would expect a Meestrol orDeepSeek to be here, but they also don’thave the funding to justify it. It iscool to see though. They now have MCP as well as Goose, which is an open-sourceaccessible AI agent that goes beyondcode suggestions. Interesting. AndAgents MD they also own. Veryinteresting actually. I did not realizethat Agents MD was big enough to go to afoundation like this, but that does makesense. I’ve seen more and more cooluses of Agents MD files. So, this iscool. Oh, interesting. Goose is byBlock. If you’re not familiar withBlock, they are a company that owns aA lot of different things. They ownSquare and Cash App. They bought Tidal,which is close to my heart, as wellas a few other things. It was founded bythe original founder of Twitter. A cool company. No issues with them. It's coolseeing them doing more open source stufflike this. Wait, no way. Yeah, I knowRoselle. She’s great. That's so cool. Ihad no idea she was at Block now. Thatmakes a ton of sense. Good for her. Goodstuff. Local AI agents autonomouslyengineering tasks seamlessly. This is agood set of things. It's cool to seesomething like this forming. Now, it's aweird set of projects. You have aMarkdown file standard. You have anactual big open-source AI agent toolkitand then MCP, which is a protocol for interfacing with agentic stuff. Veryinteresting set of things, especiallywhen you look at the set of members thatthey have. The block membership makesmuch more sense now. It seems like theykind of want to pawn Goose off there,but also the block founders have alwaysbeen very pro-open source things.So, it does make a lot of sense there.Back to the article.Oh, they have a whole fancy littleThe timeline of the history of MCP here. Early adoption from the cursor of the AItools back in December to February. MCPwas presented at the AI engineeringsummit. There were 2,000 public MCPservers at the time. A second spec cameout, which included enterprise off HTTPstreaming multimodal content, but alsothe enterprise solution they pickedwas lackluster at best. It’s myunderstanding that they used an existingHTTP standard for O that had neverreally been implemented by anyone. Thatkind of sucks, and as such it was notsomething many people did. It’s alsofunny that a thousand people registered for the MCP night in San Francisco,because that’s happening again tonightand I will be there. That’s why I’mending the stream early. At this point, theyhad 4,000 public MCP servers. Then, in Juneto July, the third spec came out withstructured outputs, which are essentialto have the MCP actually output thingsinto consistent shape, along withinteractive prompts as well as hardened security.They also introduced the MCP directory,and 2,000 people took part in the agentshackathon. Then, enterprise MCP hostingstarted happening on Google Cloud, Azure. And Cloudflare. MCP DevSummit sold outin London. This is also around when westarted to see pushbackon the idea that MCP is a thing that isjust how models will do applicationdevelopment and application-style work,which makes less and less sense. And that’swhen I did the videos about theCloudflare article about turning MCPsback into code and then letting theagents write code to execute instead.Watch those videos if you haven't andyou want to better understand the thingsI don’t like about MCP, as well as how wecan take advantage of the standard moreeffectively. This is interesting, though. I also don’t think the growth is thatimpressive. From the first twomonths, they got 2,000 servers, and thenover the next 10 months, they got another8,000. That’s relatively linear growth.I'm once again going to run a poll. Doyou use MCP? Yes, a lot. Greater thanfive servers? Yes, a bit less than fiveservers. Not really. Nope. I'll changeit to "tried but churned." I should haveput a "just contact 7" option. I'll doanother poll after for that. Yeah, I gaveup on Context 7, which is the only MCP Iwas trying to use after the last poll. Chat. According to my audience, 56% ofpeople don't use MCP, 16% tried it andchurned, 25% use it a bit, and 4% use ita 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 justhave not found many use cases for MCP inmy day-to-day usage. Yeah, the use caseI see for it, funny enough, is muchstronger locally. I've yet to find manyMCPs that are useful if you run themremotely. Like there's the Mac MCP thatlets you control your Mac. Of course,Super Memory are the ones doing it. This Mac MCP is pretty cool. The pointof it is that you can control featureson your Mac via MCP. There's a bunch ofTypeScript code written to do things onyour Mac exposed over a local serverthat your AI agents can call and dothings like send texts or update yournotes, or touch your contacts. And here youcan see in Daria's example, he tells itto send a message: "This is a test" tosomebody. And now it's running behindthe scenes on his Mac doing the thingsthat he requested, which is pretty cool. But those are the use cases that I thinkare the most valid right now due to thestrangeness of how the protocol is. Implemented. There's also theAbleton MCP. If you're a music nerd,it's pretty cool. Ableton MCP is anothervery fun one. It's for controlling theAbleton Live digital audio workstation.It's a very common tool for producingmusic. Ableton MCP lets you modifyyour project in Ableton using the Bottlecontext protocol. But once again, sinceyou have to run a full dedicated serverfor this to work, you have to set up theSmiththery CLI, install this MCP, andspecify the client as Claude. Youprobably have to install UV as well,because it's all a Python server andthen spin it up and run it locally in the background. MCP can'texecute things that it doesn't have. Itis effectively an HTTP server that isexposing specific behaviors to a modeland allowing it to stay connected tothis server so it can do things, getdata, and make changes. Another notablething about these examples is that none ofthem require complex auth stories becauseOAuth is one of the most annoying nightmaresacross all of these things. I don't evenknow if you can auth an MCP server andcursor. Oh, fun. You have to hardcodeEnvironment in the definition that youpaste in for the project. Fascinating. Yeah, it's cool. They do support it now.You also have to manually toggle thefeatures that you do and don't want itto have access to on and off. Fun. Youcouldn't tell; I don't use MCP and cursora whole bunch because I don't use MCP awhole bunch. You can also return imagesas long as they're base 64. Fun, man.What an interesting standard. Back tothe article because there are more importantpieces here. We're continuing to investin MCP's growth. Clouds are directedwith over 75 connectors which arepowered by MCP, and they recentlylaunched the tool search andprogrammatic tool calling capabilitiesin their API. I like that they wordedthis carefully so they don't have tocall it the tool search tool as theyhave been constantly. They did toolsearch and programmatic tool callingcapabilities which helps with productionscale MCP deployments, handling thousandsof tools efficiently and reducinglatency and complex agent workflows. Yeah, that's a big step. It feels likeAnthropic recognizes that improving thestandard is not necessarily within theirThe area of expertise, and instead of tryingto patch it again and again, it is muchmore efficient for them to hand it offto a better open standards committeethat can be designed. I know design bycommittee is terrible, but for thingsthat are standards that are going to beadopted by many, it makes a lot moresense. Make the standard external so itcan be iterated on, developed, and ownedsafely, openly, and externally, andcontinue building the new features ontop. It definitely seems like theirrecent interest has less been inchanging how MCP works and more inbuilding things around it. Like if we goback to the timeline, they got earlyadoption, they presented it, theyreleased the second spec which includedthe O stuff. Then in June and July, theydid the third spec with structuredoutputs, interactive prompts, and hardenedsecurity. And since then, a lot of thenew stuff isn't really part of the spec. Like discovery isn't really part of thespec for MCP. Async tasks are, but aremore like an implementation detail onthe thing consuming. And agenticsampling, I don't even know what theymean by that. Interesting. Allows for the server to reverse requestback to the model to do a thing. So ifyou're using Opus 4.5 and it uses theMCP for doing a queryagainst your database to get some data,and then your code realizes because youwrote the code this way that it has tohave the right SQL to generate, it canforward back that request to the modelto generate the SQL string that it hasto run. I've never seen anyone usingthis. This is really interesting too. There’s a hints section which is anarray of arbitrary JSON objects by thelooks of it, but they’re separately thisintelligence priority and speed prioritythings that are top-level keys, whichmeans that they’re probably part of thespec. Yeah, you can see that the chaos that isthis diagram. The server initializes thesampling with the sampling createmessage. A human can review thatrequest. They’re presented the requestfor approval. They could request tomodify or just approve it, and then itforwards that request to the LLM,returns a generation. You can review theresponse. The response is presented toThe human for approval, and if theyapprove it, it gets sent back to theclient, which then sends it back to theserver to continue. When I say it feels like all of this ischaotically over-engineered,do you understand? It is also kind of chaotic that the client, in this case, is something like ifyou use the Cloudi desktop app or if youuse Cursor, the server is sending arequest to the client that the user thenapproves so the client can then resendit to the LLM. This weird graph of whereeverything lands is just strange. It canalso request images back and audio. Does this feel like a standard looking for aproblem still? Like who's going toimplement this? What tools exist rightnow where if I wrote my MCP server torequest an image back to the client,that will handle that properly? I justdon't think there's going to be many. Oh, and then, the standards arecost priority, speed priority,and intelligence priority. Costpriority: higher values prefer cheapermodels. Speed: higher prefers faster. Andintelligence: higher prefers more. Capable. And the values, by the way, arebetween zero and one. Great. Then thehints allow servers to suggest specificmodels or model families. Again, theyhad Cloud 3 Sonnet and then Claude as afallback.Yeah, interesting. So, yeah, that is anew part of the standard. It is not oneI’ve heard anyone talking about, butthey did add something. But it doesgenerally feel like the standard in thespec has less been Anthropic’s focus andmore the things they’re building aroundit, which makes the punting of it off asthey build new tools on top of it makemore and more sense. It is, also, tobe fair, a really good faith play to say we’re not just building MCP for our ownuse cases and tricking you all to buildon top of it. They do clearly want MCPto be a winning open standard. I have noreason to believe that any of their wayof doing this is malicious. I still havemy skepticism around MCP, but I don’tthink they’re doing this because they’redumb or evil. I think they do genuinelybelieve this is an important thing toexist and that it is the best chance wehave at an open standard for AI to dothings on real systems. Still kind ofIt's crazy that they're making this so open,and cloud code is still closed source. They've sent as many DMCAs as theyhave to GitHub about people who sharedthe source maps for cloud code when theyleaked them in their own package whenthey deployed it. Just saying.The Linux Foundation and the Agentic AIFoundation. The Linux Foundation is anonprofit organization dedicated tofostering the growth of sustainable opensource ecosystems through neutralstewardship, community building, andshared infrastructure. It has decades ofexperience stewarding the most criticaland globally significant open sourceprojects, including Linux, Kubernetes,Node.js, and PyTorch. Yeah, Node is technically part of it too, which isnuts. Importantly, the Linux Foundationhas a proven track record infacilitating open collaboration andmaintaining vendor neutrality. Yes,absolutely yes.Node runs just as welleverywhere, which is kind of crazy ifyou think about it.Like you all know how much people complain about Next.jsbeing easiest on Vercel. Node workseverywhere other than Cloudflare, kind of,but you get the point. Very much supported standard that worksequally well pretty much everywhere.It's really cool that that's proven.Same with the Linux kernel. I would imaginea significant portion, if not themajority of the money that Microsoftmakes on Azure right now is goingthrough Linux systems. Agentic AIFoundation is a directed fund under theLinux Foundation, co-founded by Anthropic,Block, and OpenAI, with support from allthe other companies I mentioned before. The AIF aims to ensure agentic AIevolves transparently, collaboratively,and in the public interest throughstrategic investment, communitybuilding, and shared development of openstandards. The idea of agentic AIevolving transparently by making MCP anopen standard is an interesting thing. Iwould imagine that transparent evolutionof AI and agentic work would involvemore transparency around how the modelsare being trained, which none of thelabs are willing to do. Some deeplyskeptical part of me wants to see thisas an attempt to look more transparentwhen they are more than ever closing offtheir research and work so people can'tcopy their details. But I do really wantTo read this in good faith, there'sno reason not to at this point, at least. I'll be keeping my eye on it. And nowthe donating MCP section. Anthropicis donating the Model Context Protocol to theLinux Foundation's new agentic AIfoundation, where it will join Goose byBlock and Agents MD by OpenAI asfounding projects. I didn't know Agentswas OpenAI. That does make sense. Bringing these and future projects underthe AIF will foster innovation acrossthe agentic AI ecosystem, as well asensuring that these foundationaltechnologies remain neutral, open, andcommunity-driven. The MCP's governancemodel will remain unchanged. Theproject's maintainers will continue toprioritize community input andtransparent decision-making. There is alot of work on this. I've seen so manyissues that are hundreds upon hundredsof replies deep, as well as their sharedDiscord server. They are trying to make it better. I'vehad a lot of people, like more so thanalmost anything I've complained about,I've had a lot of people who are part ofthe MCP spec and standards committeereplying and reaching out, trying toFigure things out, especially the offthings. It seems like the off stuff's atotal disaster. The future of MCP opensource software is essential forbuilding a secure and innovativeecosystem for agentic AI. Today'sdonation to the Linux Foundationdemonstrates our commitment to ensuringMCP remains a neutral open standard. We're excited to continue contributingto MCP and other agentic AI projectsthrough the AIF. Yeah, this is cool. Ifthey want to prove that it's an openstandard, they did it right. But it alsokind of seems like Infopa can only haveone open source project at a time. Andnow that they have Bun as their big opensource project, MCP has been moved away. OpenAI's article about this foundationhas even more interesting stuff. I'mgoing to read this a little too. Similarblurb here. It is cool that they'recalling out that they're working withAnthropic and Linux Foundation, allof them. I've never seen OpenAI call outAnthropic as a supporter and Anthropiccall out OpenAI in matching articles thesame day. Weird to see this muchalignment and genuine collaboration onthe standards here. Why open standards? Matter? Developers are rapidly adoptingAI to build more capable agentic systems,from coding assistance to workflowautomation and customer service agents.In 2025, these systems have begun tomove from prototypes into tools thathandle real work in business andconsumer settings. We believe thetransition from experimental agents toreal-world systems will best work atscale if there are open standards thathelp make them interoperable, like astandard for a markdown file. Openstandards make agents safer, easier tobuild, and more portable across toolsand platforms. And they help prevent theecosystem from fragmenting as the newcategory matures. As more agents beginhandling real responsibility, the costof fragmentation increases. Withoutcommon conventions and neutralgovernance, agent development risksdiverging into incompatible silos thatlimit portability, safety, and progress. You were around for the early web daysbetween the 15 different attemptsto do dynamic web apps. I see whyjumping on this early is a thing they'reall focused on. Agents SDK isn't opensource, 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 protocolis cited here as proof that they reallywant to contribute to these buildingblocks and open standards, and codec CLIwas a pretty ballsy move. When theydropped that as an open-source project,I was surprised. I did not see thatcoming, especially after how strictlycloud code has stayed closed source.It's a bold move on their part. Andthen they did the GPTO OSS models, whichbelieve it or not, I'm still usingGPT-120 bill quite a bit. It runs reallywell on my laptop, too. Is the apps SDKopen source as well? There's no way. These are just examples. That's what Ithought. The apps SDK is not opensource. Also, to give Enthropic a littlemore grace here, the apps SDK and theidea of apps in ChatGPT was so hypedwhen they announced it, and I haven'tseen anyone use it for anything — nota single thing. Apparently, the UIportion is open source, but yeah, noone's using it. I don't care. They'vebeen early adopters and corecontributors to MCP, incorporating it asa foundation for connectors and apps inChat GPT. Just last week, they announceda collaboration with Enthropic and MCP UI toextend the apps SDK to all MCPdevelopers through MCP apps. Oh, great. They’re combining the two standards Idon't like.And they donated Agents MD.Agents MD has been adopted by more than60,000 open source projects and agentframeworks including AMP, Codeex,Cursor, Devon, Factory, Gemini CLI,GitHub, Copilot, Jules, and VS Code,among others. Jules mentioned that’sGoogle's agentic coding tool that nobodyknows about. There’s also a bunch ofother things they’re not calling outhere. Like I know that code rabbitand Grapile both use Agents MD very wellas well, which is super helpful duringcode review steps. It’s cool that wedon’t have the problem that we have witha JavaScript project like the classicconfig hell. Look at allthese dot files. The .gitignore and.vmrc, .eslintrc, prettier ignore, prettier rc,components.json, nip.json, agentmd; allof these things. It is pretty cool that theAgents MD is now just one file that workswith everything, and we’re not going tohave as bad of config hell. There was a bit where you had to have thecursor MD cursor rules agent MD winds,surf's equivalent and like 15 otherthings. Agents MD has been standardizedenough that everyone uses it. Thatsaid, different tools benefit fromdifferent types of instructions, which itvaries. Still overall, a very good thing,especially if more and more of thesetools adjust to follow the agent MDstandard better. This seems cool. Ireally like this type of collaborationand alignment. I think it leads towardsbetter things happening. I really likethe idea of the Agentic AI Foundation, where if a thing becomes popular enough,that is meant to be open, the LinuxFoundation can now own it as part of theAif. This makes a ton of sense. I wish Ihad more negative to say because it’s fun,but this is a good thing. This is thebest future for MCP. This is a reallygood thing for all of these standardsand it makes the future of Agentic AItooling much, much safer and potentiallymuch better. Curious what y'all thinkthough. Am I way too excited about thisor is this just another weird MCP thing?, Anthropic Gave MCP to the Linux Foundation, YouTube, 01kc6sb7j18g6spccq8t9ve5ge.