Really looking forward to more videos diving into uses of this! Would love to see how this might allow using a knowledge graph type system for an agent to continually improve it's knowledge based on other conversations. For example, being able to tailor a coding assistant over time to know which libraries you use, understand your preferences for architecting systems, knowing what other systems you've already created, etc.
I am struggling to see the difference with agentic swarm patterns. Agency-Swarm has provided this separation of concerns, which is abstracted into agents and their tools. The difference here, perhaps, is just new terminology for microservices and a natural language API. Don't get me wrong; swarms of agents are positioned in the same frame, which is an anthropomorphic approach to effectively object orientation on the service layer, loose coupling, and state mutation. Perhaps instead we should describes services communicating over natural language with structural responses that allow for more specific data transformation. In GraphQL, the context is passed much in the same way context is shared in the swarm of agents pattern. Once we have settled on such a protocol, perhaps we can move from agents to monads and bring typed composition with lazy evaluation through free monads. It seems we choose new terms of endearment when we try to push an ideology in programming. I would love to understand where I am missing the pieces that keep LLMs and agents away from dialects (protocols) of objects, functions and services.
My take on MCP is that, while not perfect, it presents an inevitable proposition for the future of AI-data connections. Today, APIs are already an established standard, but as AI advances, a new protocol becomes necessary to integrate models with existing systems more efficiently. Despite the initial adaptation challenges, something like MCP is a natural next step, just as APIs were in the past. I believe that in the future, people will not interact directly with each company’s chatbots, but rather with their own personal AI assistants. These assistants will access information connection points that companies need to create, enabling efficient and automated data flow. This makes MCP a central piece of this evolution. Ultimately, it represents an inevitable and promising advancement for the future of AI interoperability.
This is similar to the idea that Yann Lecun and Meta have been pushing for or saying would be inevitable. That is, each person has an assistant that acts as an intermediary / firewall between them and the rest of the world. It would certainly benefit Meta greatly to be there to intercept everyone's communications, thoughts, and actions. Of course, it does also sound beneficial to the individual in many ways if implemented in a privacy-preserving way. For example, I think we need such a firewall today in social media given the endless misinformation, let alone the attention hungary algorithms. So it may in fact be something we will all rely on sooner rather than later. PS: This text was fully human generated
Nice this video is very helpful. It would be great if you did a more detailed tutorial in how to create a simple custom MCP server to understand how it all works.
This sounds extendable to the internet in general. Sites need a standardized IO functionality that's purpose built for model consumption (non-human abstraction) and efficient interoperability.
I am a big Claude fan, it’s my favorite LLM, but Claude needs integration to websearch, that functionality is sorely missing. Perhaps this api can help incorporate search.
Very good overview! It essentially sounds like a local Zapier that interfaces with your AI pipeline without having to send your credentials and data to the cloud. Or am I way off?? Lol
Not sure what C.Desktop includes, but I'm using open webui as a chat interface to access llm APIs. It has all sorts of features, it's stable and is open-source
Yeah, I totally agree about the gaffer tape bit. I'm not sure that Microsoft or Apple are going to expose their own internal APIs to external people though. Maybe this could be the thing that actually helps Linux take off a lot more.
@ “Could not attach to MCP server filesystem” I tried everything in documentation and troubleshooting and all the pre-made tools. Connected, appear in the dev settings but it is not working and it is not shown in the app ui.
Obviously, It would be great to see a new vid on integrating RAG apps and their vectorDB's with Claude for Desktop via their "MCP" !
Agreed totally.
Can't wait for next video. This MCP feels like a revolutionary and powerful idea.
Really looking forward to more videos diving into uses of this!
Would love to see how this might allow using a knowledge graph type system for an agent to continually improve it's knowledge based on other conversations. For example, being able to tailor a coding assistant over time to know which libraries you use, understand your preferences for architecting systems, knowing what other systems you've already created, etc.
This feels like ODBC functionally back in the day. God I'm old.
EDIT: corrected grammar
I hate the fact that we STILL use this in ancient excel files instead of dedicated software implementations
@AngeloXification ikr?
So basically it is like OpenAI plugins but then open-sourcing it as well. Nice! No need to write & deploy my own LLM API wrappers anymore!
Thank You. Very Timely !!
I am struggling to see the difference with agentic swarm patterns. Agency-Swarm has provided this separation of concerns, which is abstracted into agents and their tools. The difference here, perhaps, is just new terminology for microservices and a natural language API. Don't get me wrong; swarms of agents are positioned in the same frame, which is an anthropomorphic approach to effectively object orientation on the service layer, loose coupling, and state mutation. Perhaps instead we should describes services communicating over natural language with structural responses that allow for more specific data transformation. In GraphQL, the context is passed much in the same way context is shared in the swarm of agents pattern. Once we have settled on such a protocol, perhaps we can move from agents to monads and bring typed composition with lazy evaluation through free monads. It seems we choose new terms of endearment when we try to push an ideology in programming. I would love to understand where I am missing the pieces that keep LLMs and agents away from dialects (protocols) of objects, functions and services.
Go do that then
@ i hear you “talk less just do it” …ok then.
thanks for this impressing tutorial, yeah pls dive into this, this look like a future standard.
Thank you Sam, for sharing these updates
Big thanks!
Thanks for the quick and easy showcasing! 🙏🏻🤗🔥 Things are getting interesting!
Great video Sam
My take on MCP is that, while not perfect, it presents an inevitable proposition for the future of AI-data connections. Today, APIs are already an established standard, but as AI advances, a new protocol becomes necessary to integrate models with existing systems more efficiently. Despite the initial adaptation challenges, something like MCP is a natural next step, just as APIs were in the past.
I believe that in the future, people will not interact directly with each company’s chatbots, but rather with their own personal AI assistants. These assistants will access information connection points that companies need to create, enabling efficient and automated data flow. This makes MCP a central piece of this evolution. Ultimately, it represents an inevitable and promising advancement for the future of AI interoperability.
yeah agree this makes sense in many ways.
this post sounds really AI generated
@@absta1995 I structured my thoughts and just gave it to people to organize and write. hehehehe
This is similar to the idea that Yann Lecun and Meta have been pushing for or saying would be inevitable. That is, each person has an assistant that acts as an intermediary / firewall between them and the rest of the world. It would certainly benefit Meta greatly to be there to intercept everyone's communications, thoughts, and actions.
Of course, it does also sound beneficial to the individual in many ways if implemented in a privacy-preserving way. For example, I think we need such a firewall today in social media given the endless misinformation, let alone the attention hungary algorithms. So it may in fact be something we will all rely on sooner rather than later.
PS: This text was fully human generated
In my case MCP works great on MAC version of Claude app with PC I had some issus:) Idea is great but needs update:)
Nice this video is very helpful. It would be great if you did a more detailed tutorial in how to create a simple custom MCP server to understand how it all works.
yeah planning on doing this
I don’t know why RUclips recommended this to me. But this sounds like what I want to make a career out of
Now, this Anthropic "MCP" sounds like it has / will have similar capabilities to the "Tron" "MCP" ;)
lol, that was my first thought - the MCP cone.
Thanks 👍
This sounds extendable to the internet in general. Sites need a standardized IO functionality that's purpose built for model consumption (non-human abstraction) and efficient interoperability.
yeah it probably needs Auth added to it as well
Thanks Sam :)
Do you have a video covering how to launch a website if you make one in vs code
Is this basically function calling or GTP actions?
more two way tool use
I'm planning to make a robot to click that [accept] button automatically!🤣🤣🤣
I am a big Claude fan, it’s my favorite LLM, but Claude needs integration to websearch, that functionality is sorely missing. Perhaps this api can help incorporate search.
Yes, you can totally use it with Brave Search.
Very good overview! It essentially sounds like a local Zapier that interfaces with your AI pipeline without having to send your credentials and data to the cloud. Or am I way off?? Lol
do you think Windsurf already made use of this or are they on Agentic route?
honestly not sure I haven't had a chance to check out their product
genius, it's kinda similar to what ollama is doing, an integrated way of interating with base llms
Is there any open host that could replace Claude Desktop? I can't find anything.
Not sure what C.Desktop includes, but I'm using open webui as a chat interface to access llm APIs. It has all sorts of features, it's stable and is open-source
currently not but my guess is there will be a VSCode extension pretty quickly and probably others as well.
How is this different from function calling?
Agent Pilot will support it soon
I can imagine a combination of puppeteer and the file system could almost entirely replace perplexity.
This is just like SwarmZero’s idea for an agentic protocol
Is this a step towards AI optimised operating systems? Currently I feel like we are gafa taping AI onto systems not designed for it.
Yeah, I totally agree about the gaffer tape bit. I'm not sure that Microsoft or Apple are going to expose their own internal APIs to external people though. Maybe this could be the thing that actually helps Linux take off a lot more.
Siri+LLMfarm+pythonista+shortcut
same stack
Sadly all new features don’t work well on windows 😔
what problems are you seeing ?
@ “Could not attach to MCP server filesystem”
I tried everything in documentation and troubleshooting and all the pre-made tools.
Connected, appear in the dev settings but it is not working and it is not shown in the app ui.
Agi
Snowflake integration with Claude MCP to enable intelligent data pipelines and AI workflows 🤫
It's a bit ironic, that this video is voiced by AI.
but I am an AGI AI
@@samwitteveenai Oh... then I won't tell anyone ^^