Claude, TypingMind, AMP & MCP Servers: The Future Dev

9,594 words · about 38 min to read

LLMs accelerate non-traditional developers into software creation

0:00 I get a ribbing from my friends because I say I'm not a developer. If you were already using a web tool like Claude or ChattyPT, I think one of the ones I would look at is typing mind. If I go back to the chat and I say write like Popey, this has a prompt in it that says write a blog post the way Allan would write. You're not locked to one of the providers. I could start the chat and say actually I want to do this with Gemini or I want to do this with Chat GPT. People are trying to learn the best practices for themselves probably way too much in isolation.

0:31 Sometimes you configure an MCP in your agent and then it goes away and so it can be a bit dangerous having this tool hanging around in your MCP.json file that you think it could be that that's actually a liability.

0:45 What would you say is one of the next big things that we should be thinking?

0:48 I think I'd say Before we jump into this episode, I wanted to let you know that this podcast is for developers building with AI at the core. So whether that's exploring the latest tools, the workflows, or the best practices, this podcast's for you. A really quick ask, 90% of people who are listening to this haven't yet subscribed. So if this content has helped you build smarter, hit that subscribe button and maybe a like. All right, back to the episode. Hello and welcome to another episode of the AI Native Dev. And last week we had a little bit of a a 101 with Max on AI agents. And today we're going to continue that practical theme and continue talking about various agents, how you can use agents, but also how we can extend the capabilities of agents through MCP servers. and not just talk about them, but we're actually going to see and have a play with uh some of Alan Pope's uh favorite MCP servers and and

1:51 servers that he's played with in the past and talk about the tools uh that provide the capabilities that those MCP servers uh offer. So, yeah, I mentioned Alan, Mr. Alan Pope. Alan, very very welcome to the AI Native Dev podcast. How are you?

2:04 Thanks very much. Yeah, it's great to be here. I've been watching for a while and now I get to be on it, which is lovely.

2:10 Long time long time listener, first- time caller, right? Yes, exactly.

2:14 So, Alan, tell us a little bit about who Allan was before we met maybe and then and then today

2:20 I think I I get a ribbing from my friends because uh I say I'm not a developer. Uh I didn't do I didn't study software engineering. I've been a Devril and community manager for a long time and I can write some code. But I found in the last few years, all of the different developments with LLMs have accelerated me from the ideas that I have uh turning them into actual usable pieces of software and actually sometimes even publishing them on the internet for people to you know critique, review and use and contribute to. And so I think um it's been a real accelerant just for me personally using uh LLMs because I would often find it difficult to get started, not know what tools to use, not know what frameworks to use, uh not understand everything that I was doing and this has just made life a little bit happier for me cuz I can finish things.

3:19 Yeah. Amazing. And and when we when we first met, we we grabbed a coffee in a in a in a local Starbucks and we just

Technical specifications bridge human intent and LLM execution

3:28 kind of like geeked out on all things specifications and and I guess what what was the what was the drive that kind of made you lean more into specifications? What was the problem that made you think, do you know what, we need to we need to grow this out into a larger spec versus versus more of a prompt style

1 Highlight · Using technical specs to guide LLM code generation

3:48 approach?

3:48 Yeah. So I think I'm already my mind was already in the space of I want to describe what I want because I felt like if I can [clears throat] describe what I want uh at some point I will have the time and capabilities to turn that into code right but I what I need to do first is write it all up in English in a way that I and other people can understand. So, I would often run these I wouldn't call them specs. They were like design documents and I'd run them past uh [clears throat] other friends who were software developers and sometimes I'd end up nerd sniping them into doing it for me. And sometimes I'd just get their critique, their review, and they'd say, "No, that will never work." Or, "Yes, that's a good idea. Someone should do that." And it would go filed away in the big list of projects that Alan would like to finish one day. Uh and then when the whole specbas based development started uh like gaining traction, I felt like this was my time because there are these uh documents that I've written that I fully understand. [clears throat]

4:50 But communicating that and conveying that into you know a uh a software developer's head is hard. Um, but [clears throat] conveying that to an LLM that understands what I want as far as they can understand and can turn that into how to make that a thing that I can run and with a spec that is more technical and not so much a design document, but a technical specification of okay, we're going to use Python, we're going to use UV to do dependencies, we're going to have a a particular database for the back end, and this is what the API should look like. That's the bit that I was missing.

5:26 And with the specs I can identify best practices and also uh [clears throat] identify how I implement those best practices. And so it's kind of that's the bit that's missing. And it there was quite a learning exercise in discovering

5:39 all of the the ways in which you use these tools and agents and and specs. And I think I'm quite keen on sharing what I do because I feel like there's a gap and people don't often know how to get from chatting to an LLM to creating a finished project. And that's that's sometimes because they just don't know what they don't know. And so I'm I'm very keen on sharing that kind of information.

6:03 Interesting. And a lot of what you were saying is really guidance. It's not necessarily just what the thing you want to build is. It's a lot of guidance in terms of how you want the agent to go about building that as well. The It was funny actually. We we chatted in the office the other day and uh you you were we were talking about tools and tool usage and and one of the one of the things that we talked about is how you go about finding and identifying those tools. And we're going to talk about a ton of tools here today. We're going to talk about we're going to talk about claw gemini CLI uh uh typing mind uh we're going to maybe talk a little bit about various MCP servers that do things like sequential thinking and and uh fire crawl and and maybe Tesla YouTube uh access and all those types of things.

6:51 But Alan, how do you kind of identify, you know, where you want to uh grab these tools from? what what what kind of what's the thing that makes motivates you to try them out and and and learn about them?

7:06 So, initially it was I just needed the the thing to get going and so there there are a few uh tools that I've used that got me got me started and now I have an understanding of the landscape. I'm trying to find new things that can make me go faster or replace manual processes. So, for example, one of the ones you mentioned is a YouTube transcription downloader. So, go and get the transcription of a video because there's a lot of words in there and you could use those words in the creation of other content like social posts and blog posts and that kind of stuff. And as part of Devril, that's something I would do. Now, I'd already written a [clears throat] little shell script that uses YTDLP, a popular like YouTube downloader tool, but [clears throat] that was manual. I had to like run YTDLP and the URL. Whereas what I was thinking would be better is if there was an MCP that could do that for me while I'm in the flow of creating a blog post and I'm in the zone and I don't have to jump out to use some terribly written shell

8:07 script from me or uh use some third party tool. I can have it all integrated. And so having a conversation with an LLM where it can branch off and go, okay, I'll now I'll go and get those transcriptions and then now I'll do the summary and then let's figure out what keywords people are looking for and do all those things but outsource them to the tools that can do those things best. That's that's the the part that I find interesting. The the tricky part is where do you discover these things? And sometimes it's it's simply word of mouth. I sit having coffee with a friend and they tell me the things they've discovered over the weekend and that's not scalable. It's great for me and great for him but it's not scalable and I think we do need better ways to share uh the tools and settings even like

8:55 [snorts]

8:55 uh the the little configuration settings for each of your agents those kind of things and then you sit there and read through all of the documentation for everything you know and ain't nobody got time for that. So, I'd way rather have recommendations from other people and rated um tools and MCP servers that that I can

9:14 make a qualitative decision about whether I'm going to use that or not.

9:18 Yeah. Yeah. Cool. Let's talk about

TypingMind provides a configurable multi-model web agent platform

9:20 agents a little bit now and jump into uh I guess which agents you had potentially talked about word of mouth and and and recommending these agents. For someone who's never really used, let's say, a coding agent before, we have to first, I guess, recognize there are different styles of agents, a an agent is just something that can go away and do something very autonomously. Uh perform various tasks. Uh maybe maybe even call out to sub agents and other agents to perform those tasks. Uh it doesn't necessarily say anything about the UI as to how you interact with that. very often chat based uh sometimes in a UI that is web oriented sometimes others.

9:59 Where where would how would you recommend to people to try and use an AI agent first? Is it up to them? Is it the best tool on the market? Where would you go? You know, this is one of those tricky things where it's changing all the time and I know a lot of people uh just use chat GPT and they just chat to it in a web page or in the app and they'll ask it to create something and it will do that. And certainly over time all of the agents that are available, the popular ones have improved dramatically. They It used to be you'd get into an angry argument with an agent trying to get them to do what you want and get them to understand what you want. Um and eventually you'd fill up their context window and um it would forget what you originally asked them for. And things have got a lot better now where many of the web- based agents have code editors built in and that some of them have got a place where you can run and test the code directly in the web page. And so [snorts] my experience

11:00 from early use of chat GPT to write code or early use of uh claude in a web page to use code is probably very different now um because people could just open a web page and they've got a code editor right there. Um so I think my experience might be colored by the the fact that I've been doing this for a while. And by a while I mean like a year or more like that's still not very long. Um, so I think if you were already using uh [clears throat] a web tool like Claude uh or chatpt I think one of the ones I would look at is uh typing mind. Now, type in mind is interesting because it's same web type UI, but it's super configurable in that you can connect it to external tools to do things like download the YouTube um uh thumbnails or the uh transcription or get it to go and search online for more information or connect directly to GitHub or whatever

12:02 the things you need to do as part of your development to do research and and

2 Highlight · Using TypingMind knowledge bases to replicate writing style

12:06 software development. But the nice thing is it has this web UI, but it is super configurable. And so, um, if you look at the example I've got here, you can create these things called agents. Now, these are, uh, you know, when you think of the term agent, it's a bit of an overloaded term. And in here, it means like a personality of an agent. So, it's it's prepared with a a prompt, and it may have access to some knowledge that you've uploaded. So, for example, here's a very simple example that I've used.

12:37 this one here called write like popy now all I've done is in typing mind I've updated a knowledge base which over here on the left hand side and in the knowledge base you can upload documents that uh your agents in typing mind will have access to and so down here I've got one called popy blogs and I've had a blog since about 1997 or something and I grabbed all the source for all of the articles that I've ever written and put them in here so that I can have typing mind help me write an article that sounds like I wrote it. Now I realize that sounds like I'm just, you know, not ever going to write any any blogs again because I'm just going to leave it up to Typing Mind to do it, but I did it as a test. I haven't actually used it an awful lot. I've used it a couple of times, but it's quite remarkable that once it knows your style of writing and your tone of voice and the way that you write, uh it's very um

13:37 uh able to create new content that looks like you wrote it. It's it's quite surprising. I can I can give you a little demo. If I uh go back to the chat and I say write like Popey, this has a prompt in it that says write a blog post, you know, the way Allan would write. And down here you can give it access to those MCP servers that we'll talk about a little bit more a bit later. [snorts] But you can also give it access to um the knowledge [clears throat] base. And if I say, you know, um, right, uh, oh, right, I can't type, write, uh, blog about new developments, MCP servers

14:21 in the command line. So you the other thing you'll notice while I'm while it's chugging away, you'll see down here it says cla. Um, one of the nice thing about uh, typing mind is you're not locked to one of the providers. I could start the chat and say actually I want to do this with Gemini or I want to do this with chat GPT and you can even fork conversations partway through and change from one model to another. So this is quite an advantage over just sitting in the chat GPT web UI or the clawed web UI because I I'm you know claudes don't give you access to chat GPT and vice versa. And so there's some interesting options in here. Now, you'll notice here it's gone and found

15:08 [clears throat]

15:08 uh a blog post that I wrote some time ago called Command line only laptop, which is a ThinkPad that I had a while ago. Um, and then it like reads a few more things, some other articles that I've written. Um, and down here, I mean, this is exactly the style of content I would write. I put a TLDDR at the top and then I'd write far too many words. And so it it's got that style exactly right. Um and you know subheadings like what on earth is MCP?

15:37 That kind of stuff. That's the style of writing that I have. So the [snorts] thing I like about typing mind is it's doing a lot of things all within one user interface. It's able to connect to different uh tools. It's able to uh have [clears throat] knowledge uploaded to it. Uh not just blog posts like documentation or existing code. Um and I can access this from anywhere. uh because I I've signed up and I can access this on my mobile phone and have a conversation and actually start a conversation and some software being developed and then walk away and continue somewhere else. One of the other things that it has which is really cool [snorts] is one of the uh MCPs, one of the tools that it has access to is called file system. And this is handy because I'm on a web page, but I can give this tool access to specific directories on my machine and say write the code directly on my machine. And that gets away from what we used to do, which is like have a conversation with an LLM and copy and paste the code out of the conversation. With this, it just

16:40 goes directly to that folder and will start creating files. And when you go to that folder, you'll see the the software we created. And then later on, I could use the GitHub MCP to have it actually create a repository on GitHub and upload the whole thing and upload the readme and even make the readme sound like me and in the style of my other readme. So, it's I find this is a good um tool type in mind, a good agent for doing a lot of different things in one place without you having to switch between, you know, all kinds of different uh tools and utilities. And if you prefer the web UI, this is great for that. And it's amazing because it feel it feels quite generic in the sense of it's not necessar specifically for coding or specifically for writing blogs. But I guess that's the power of MCPs uh an MCP servers to be able to connect you with those specific uh tasks that you might uh that you might want the agent to branch out and call an MCP directly for. Um I'd love to I'd love to maybe do some testing here, Alan. see if uh see if you

17:42 could write 10 10 different blogs or so and uh five of which you wrote manually, five of which uh you use this for and we can see maybe if uh if if folks who know your style of writing can guess which ones are from you, which ones are from AI? Uh do do you have to do much editing on on these blogs? Yeah, I I will always edit because they sometimes still do get a little bit excitable and uh you know use phrases that I would never use or use

18:08 [clears throat and snorts]

18:08 um terms superlatives that I would never use but it's surprisingly good in terms of is it convincing. Um, I also used this to create uh webinar abstracts um in a past job. And the way I did that was I looked at the potential person we were going to talk to and I scraped their LinkedIn of every post they've made recently, their social media. I then went to the open-source project that they maintain and I scraped the last five uh meeting notes for their open source project that that they talk about. And I put that all in there and I said, "Give me a webinar synopsis that this person could talk authoritatively on topics that they clearly like." um and then gave them the webinar [clears throat and snorts] uh format, synopsis format that we use in the company and it came up with an abstract that was pretty much spoton and I sent

19:10 it to uh my boss and they were like, "Yep, that's great. That's exactly what we're after." And I sent it to the person I was going to be discussing this with and they said, "Yep, I can talk about all of that." I was like, "Well, there we go.

19:21 We totally should have done that for this podcast, Alan. Let let's switch.

19:27 Yeah, maybe. Maybe. Who knows? We'll never know. Let's um let's talk a little

Terminal coding agents streamline rapid local prototyping

19:32 bit about uh a more specific tool now cuz like like we said, this is this can be used for many many things uh including those synopsis. And I guess you could use something like Claude Code for similar things and it probably won't complain. It'll do it'll do what it's told or what it's asked for. But Claude code is much much more of a developers tool in the sense of if someone wanted to to to make changes to code, create new projects, uh claude code could be a a great a great starting point. Um talk us through terminal AI agents and I I guess this really only appeared earlier this year I suppose right? Um, are you a fan of the first of all, what is a terminal UI and are you a fan of it?

20:17 Good lord, yes. Uh, uh, so my, uh, when [clears throat] Cloud Code first came out, I was, um, using VS Code and GitHub Copilot. So a familiar graphical IDE that everyone's seen and down the side you have a chat with an agent co-pilot and it you know helps you update or create projects um directly in the in the IDE with Claude uh it's a little different um but it depends how you use it like you can just uh jump into a terminal and we could uh create a project in AID and call it I don't know um [snorts] project five. This is my fifth project that I'm doing today, I guess. And just launch Claude and ask it to create something like there's there's no IDE here. I'm not looking at any code. I'm not um within

21:18 some other tool. I am directly talking to Claude here. And at this point, I can ask it, you know, create a Uh [snorts] right now we're all criy that's uh API error [clears throat] that happens. Uh so luckily it recovered. Uh so I just guessed what to do there. This is not a prepared demo. I literally just said create a thing in a language that I I don't know uh to do a thing um that I would like to achieve. Right? I I quite like having RSS readers. Now, under the covers, obviously, this is going to be creating code in this folder and that will appear at some point. It's going to create files and I don't know, maybe I do or don't have cargo installed or I do or don't have all the relevant bits, but

22:19 it's going to start um munching away on the knowledge that it has and whatever it's found via whatever tools it has access to and it's going to start creating some code. Whether it works or not, I don't know. But [snorts] I've not had to give it an awful lot of preparation. I literally that one line is all I told it and I haven't done any prep for this. That was just a you know spur of the moment suggestion and that's quite freeing like not having this entire IDE in front of you but just having a chat with a thing that may or may not be able to create what what you envisionage right now what I'm envisaging is yeah a list of uh feeds and I click on a feed and I see the the most recent um articles and I click on an article and it expands them. I don't know if that's really what this is going to create. That's what I have in my mind. And I think that's the whole um dare we say it vibe coding of a few

23:19 months ago was [clears throat] very much like just blurt out one line and just say, "Yep, let's go. Do it, do it, do it, do it, do it." And away it goes and creates a thing or it doesn't. And it may work or it may not, right? I I don't feel I want to do that. I want to create things where I can clearly articulate what it is I want. Like I don't think it's enough to say create an RSS reader in in Rust. It might be and what kind of what kind of claude code developer are you? Are you or what kind of aentic developer are you? Are you the kind of person who short prompts and then iterates to try and say actually I would rather this than the way you did it or uh I want to you know tighten up these requirements over here or do you actually long prompt to start off with whereby that first prompt is quite

Verbose prompts and automated permissions maximize agent autonomy

3 Highlight · Evolving from short prompts to verbose initial specifications

24:07 detailed in terms of what the style of the app is maybe what stack you want it to be run on and a number of capabilities which which do you naturally fit into? I think I've evolved from and I think what a lot of people have done is they've evolved from that the short prompt to the to the the longer prompt uh simply because it can get quite exhausting uh when you you you have an idea of what you want but you didn't articulate it fully um and so the the the agent tries the best account with the information it's given um and so what I now tend to do is be very clear, very specific, almost too verbose because I I really want to steer the agent from the start.

24:57 And what I will often do is set the permissions. So these little prompts that pop up that you're seeing now where it's asking, should I have to install this thing or should I u modify these files in the current directory? I will often say yes to all of those or configure in my claude settings to allow it to edit in that folder because I actually want to give it a quite verbose prompt and then I want it to go away and leave me alone for a while so that when I come back there's something useful for me to see and I'm not I don't come back to the screen and see it's waiting asking me a silly question, right? I want it to just put as much effort in first and then ask me. And in fact, I've got a setting on uh my Claude because I quite like multitasking. I'll [clears throat] have Claude running maybe while I'm watching a film or maybe while I'm making dinner for the family and I'll have an idea and I think, "Oh, let's see if Claude can do this." And I'll type as much as I possibly can and I have a clause setting that will say out loud when it's waiting for my input.

25:59 And so it will say, "Claude is waiting for input." And so if I've got the speaker turned up or I've got my headphones in, then it will let me know that it's it's ready because it's it's dead time. If it's prompted something and I haven't noticed and it was an hour ago, then it's a whole wasted time. Whereas uh if I have it tell me it's it's waiting. Come on, Alan. What? Answer this question. Uh tell me what I

26:21 should do right and tell me what I should do next. So yeah, I I use lots of words um to try and coers it in the right direction. I'm glad you [clears throat] mentioned headphones there cuz I can just imagine the family going, "Oh, dad, I'm trying to watch a film." Do

26:34 you know what's even worse is it's my voice that it does it.

26:39 I use um I used to talk Piper on Linux. Uh and many years ago I donated my voice to an open source project and because my name is Alan, my voice model appears near the top in hugging face and so people download and use my voice for stuff and I use it as well. It's quite funny having my own voice coming out of the computer saying Claude is waiting for prompt.

27:01 So for people who have used that in the

Comparing Claude, Gemini, and AMP coding agents

27:02 past, uh this could be a quite a quite a strange episode for them who are listening to you directly as well, right? They're thinking they're their kind.

27:11 Yeah. Awesome. Um so yeah, great. And and Claude is your go-to tool these days.

27:19 So yeah, I think I think I'd say Claude is the primary one I jump to. Uh, [clears throat] and I tend to use Gemini, uh, if I've run out of tokens include. Uh, so if I've, you know, I don't really want to keep throwing money at a project. Um, and so if if it's not useful for me, um, and you've got to be tight on the budget. So sometimes I'll just switch to Gemini because I have a Gemini account, right? And it's very similar, but Gemini is seems to be a little bit more constrained on how you can prompt it.

27:52 And so I don't feel as free with the Gemini even though it has a much higher context window, larger context window. I feel I don't feel as free because I feel like it sometimes says no, I'm not going to do that. Um and certainly I see that in typing mind. The prompts that you give to Claude will work with the same prompt you give it to Gemini and it won't. Um the one other uh command line one that I've used recently is AMP which has an interesting model um where they have a free ad supported tier. Um, and I gave that a test this week and uh actually asked it to create a web-based RSS reader. Um, and it did and I was quite surprised how good it was and I didn't give it any tools. There were no MCPS connected to it. I [snorts] just used pure AMP on its own and it was very much like claude but it

28:44 Yeah, it amp is the AMP is the agent coding agent for from source graph

28:50 I believe. I think I think I'm making a huge assumption that we should we should check that afterwards. Um yeah, I'm pretty sure I'm pretty sure

Scraping GitHub to uncover real-world MCP server usage

28:57 AMP is from SourceCraft, creators of Kodi as well, which is uh which is nice.

29:02 Um so should we jump over to MCP servers?

29:05 Yeah, sure.

29:06 So MCP servers, there are registries after registries. In fact, there are registries of registries, I think, in in uh that that allow you to track all the different MCP servers. Actually, we had uh we had uh a great uh session um around MCP.Run, which is a [clears throat] a super interesting uh nice SAS version uh where that can host MCP servers and things like that as well. Um but when we when I I remember when I said to you kind of, you know, what what are the what are your favorite MCP servers? What are the MCP servers that people mostly use? Tell tell us tell us how you how you discovered what MCP servers people were mostly using.

29:42 Ah, that's a great point actually. Hi. So, we're going to jump into the second half of our conversation. But first, a quick shout out to AI Native DevCon. 2 days of hands-on sessions, deep dives, and honest conversations about what it means to build AI Native Software November 18th and 19th in New York, or join virtually from anywhere. Tickets and details at a native.io/devcon. See you there. I have my own, but it turns out other people have preferences

4 Highlight · Scraping GitHub to discover popular MCP server trends

30:10 and I I I could go and ask everyone on the internet or put a poll on social media and say, "Hey, what are your favorite MCPs?" But no, that's that's far too easy. Um, I wrote a script or I asked Claude to help me write a script which basically scraped GitHub looking for the configuration files that people list the MCPS in. So, um, MCP.json JSON is a common one that a lot of them use and it's a JSON file that lists um in a structured way all the um the MCPs and uh so you know you can see in that first slide we showed a list of some of the MCPS I'm used I mentioned a couple of them but I wonder what other people use and I got it to write a report um and this is that report. Now bear in mind this is scraping GitHub. Now, not everyone is going to um add their mcb.json to their projects and some people have private projects and some people use things other than GitHub like

31:11 GitLab or other code hosting tools and some people will have it totally in-house. So, this isn't an exhaustive list of every MCP that everyone's using. Right? So, in these reports, uh we found a whole load of uh configuration files. Uh it gave me some repositories uh that have configuration files in them. Um variables and here we go. There's a a list of some of the tools that people have in their configuration files. Now the thing that I found surprising about this um [clears throat] was that the same few tools pop up at the top of the list quite a lot. And I think this is part of what I was talking earlier is people learn from others and they copy and paste and it's almost like a cargo cult of copying and pasting stuff from one place to another. The other thing is there are repositories which are collections of other people's settings files so that people can learn. So it is

32:12 a bit wild west and there's a bit of duplication there. But you can see here there are some [snorts] of the ones that I've already mentioned like file system, being able to access a file system and sequential thinking and memory. Um, and yeah, there's there's a bunch of uh interesting ones like Brave search. So being able to use Google to search the web is one option, but there are other MCPS to be able to search the web with

32:39 KI or Brave or um I think there's probably a Bing one as well. I haven't looked, but these these are some of the ones that people are clearly using because they're showing up in their settings file. [snorts] And I'd be interested to like dig deeper into which ones people are actually using and which ones are in repos that maybe we can't see. And I I think there is an opportunity there to surface this information about MCPs that people are using because there it is a bit uh unknown and a bit daunting for someone new to know which ones to pick. Um yeah,

33:14 that's I found that

33:16 that that's really interesting not just to see not just to see which ones exist in configuration but to see which ones are actively being used in sessions to see if there's that data. There'll be interesting actually I wonder if MCP run share any of that data about usage and things like that that would be super beneficial. Um yeah and and I think it's I I really that really resonates with me the the the fact that it's people are trying to learn the best practices for themselves probably way too much in isolation and there's a real lack of the ability to share that I guess in a way that makes it easily discoverable. Um, I suspect a lot of people are probably trying to roll their own here. Uh, particularly within organizations and it's much harder obviously to add those policies outside of an organization to say to people, hey, here's some great cursor rules or or or any kind of configuration for for an agent or or something like that.

34:11 Yeah. And you also have to be really careful because sometimes people are putting

34:15 [clears throat]

34:15 um secrets in their MCP configuration. And so they actively don't want to share that not because they don't want people to know which MCP tools they're using, but they don't want to accidentally share their GitHub uh personal access token or you know some other more important uh API token. And so there is this feeling that um I shouldn't share this, but I feel like there is a space where we could have [clears throat] um a database of MCPs that are rated both by popularity, but also rated by the quality of them and whether they get software updates and whether they're still maintained. um because it's you know sometimes you configure an MCP in your uh agent and then it goes away because they've moved the repository or they're bought by somebody else or they've just given up. Um and so it can be a bit dangerous having this tool hanging around in your MCP.json file that you think well this is giving me

35:15 value. Is it is it still giving you value? it could be that that that's actually a liability and no longer um a benefit to your project and so there there does need to be some auditing of these tools and I know there are some

Orchestrating multi-step workflows using connected MCP servers

35:29 organizations already doing this uh to make sure that people aren't getting themselves uh in trouble using tools that are no longer maintained in secure and so on.

35:38 Yeah. Should we jump into some demo? Should we uh should we see what these MCP servers are all about?

35:43 Okay, let's do a demo using uh Claude. So if I open Claude, it will see that there are some uh MCP servers configured in the uh MCP.json. And if I then do /m MCP, we should be able to see the status that all of them are connected and working successfully. We probably won't need to use all of them for this little good demo. All we're going to do is say uh we've got a YouTube video over here, one that you may have seen before. Um, and I want to create a YouTube description, maybe some social posts, and a blog post. Uh, let's see if it'll do all of those. Um, [clears throat] grab the transcript of the attached YouTube video and create a description, some social posts, and an indepth blog post.

36:41 have the cups covered. Let's see how this does. So, one of the things that Claude loves doing is creating these plans, right? These a set of to-dos, plans out what it wants to do, and then uh and then goes through them.

37:00 Yeah. I find ah now you'll see there it's immediately said it's going to use a tool. So, it's figured out that YTT is the one it wants to use, and I'm happy for it to use YT. Um, so yeah, it creates the plan and then works through it. Incidentally, I mentioned AMP earlier as a another agent. It does a very similar thing, but the UI is different and I really like it. It has a a box on the right hand side and it keeps the to-do list on screen, whereas I often find with Claude, it all scrolls off and I'm not entirely sure what it's up to at any time. I'm sure I could bring it back or just scroll back. But yeah, I was I was quite pleased to see that AMP keeps it um right there on screen so I can see it at all times and see what progress it's making. That's uh that's the interesting thing with uh with with something like this. It it it provides the output directly into the into the into the terminal, right? But it doesn't necessarily doesn't necessarily unless you explicitly ask it to uh save anything to disk or or it knows it needs to perhaps for a coding file or if it wants to make a change to a coding file or something like that.

38:01 Yeah. And that's that's a one reason why I really quite like using uh typing mind because uh what typing mind will do is while you're in the web page having a session uh anything that it creates there's a copy button just right there on the screen just like with chat GPT and all the other web-based ones. But one of the nice things is you can copy it with formatting and paste it straight into a Google doc and it will come out formatted well for for uh editing [clears throat] in a Google doc which is pretty awesome. Right. So it think it's uh created some files. So now

38:33 is waiting for human response.

38:35 I got interrupted by myself there. Um which is yeah not uncommon. Um so let's look at the video description. Uh it's got the name of the person. Does this match the conversation that you had in that? Uh then the social posts are a bit tricky because social posts with all of these AI tools often like smack at being written by AI tools and so they you really need a bit of uh human interaction to do a bit of uh finessing of these but it's it gives you some ideas. It's good for ideiation and summarizing what's going on. And then the blog post um let's have a look at that. [snorts] It's comprehensive if [clears throat] if nothing else. There's a lot there. So, um,

39:20 it's a good starting point, right? Even even if there's nothing else that uh to to give

39:25 to to avoid that writer's block to have a whole bunch of thought areas.

39:29 And and to be fair, I did yolo it and say, "Write the blog post." What I should have done is say, "Give me some topics, some, you know, things, and then we could have worked on it together, iterated. But for a quick demo, I think it illustrates that like I I didn't have to manually go and get the transcript for that video." [clears throat] And I realize that's a very simple example, but there's so many of these that when you hook them all together and you have a whole bunch of uh tools working together, orchestrated by your agent, it it dramatically accelerates what you can do because you're not stuck having to go and manually do things along the way.

Automating security checks with custom vulnerability scanning MCPs

40:02 And that's that's the thing that I find is like the the thing that accelerates everything for me. I actually wrote um a an MCP server uh which is the one at the bottom of the list, Gripe. Uh so when I worked at Ankor, Gripe is a uh vulnerability uh scanning tool and um I wanted to have an MCP uh uh server that would scan the code in the repository that I'm sat in and make sure it didn't have any vulnerabilities. Now, I haven't tested this recently, so I don't know if this is going to work well. We'll give it a try and see. Right. So, for this demo, I'll create a little piece of software and then I'll see if I can get it to do a vulnerability scan. And again, I'll use Claude.

40:50 Um, and this is a fresh folder. So, it's asking me again if I want all those MCPS. Um now what I'll do is I'll try and trigger it

40:58 [snorts]

40:59 um by making it build something that will be using outofdate uh software and by forcing it to use out ofdate software maybe we'll find some vulnerabilities assuming my MCP works. So um create a hello world Django app uh app using what's the oldest version of Django that we think it'll uh be able to do [clears throat] using Django 4.0. There we go. I think that's old. That sounds old. Uh using Django 4.0 zero and use UV for dependencies.

41:41 So, all this is going to do is create a little bit of code. It probably won't do an awful lot. Um, it'll create uh a virtual environment using UV [snorts] and then install whatever Python packages are required to get this project bootstrapped. Um and then

42:00 for those for those who don't know what UV is,

42:03 so a lot of people use uh Python virtual environments so they can have isolated uh directories for each of their projects and uh most of the time people use Python VMs and there are other ways to do this but UV is a relatively new tool written in Rust and it's a lot faster than using Python. virtual environments. Um, so instead of doing uh Python pip install something, you do UV pip install something. And actually it's quite common to see MCP servers that are delivered using uvx. And so um it's the equivalent of uh node packages that use npx use uvx. And in fact that's how uh the gripe mcp server is delivered. It's written in Python [snorts] and um all you put in your MCP uh [clears throat] JSON is uvx gripe mcp and it will go and find it. Um and it's um it's

43:05 [clears throat] just a really nice way to contain all of the dependencies for your Python project in one directory.

43:11 Nice. So it's uh it's initialized it and now it's creating the project structure, the Django project structure.

43:17 So it's created a little project. Let's see if we can use the GRE MCP to scan

5 Highlight · Integrating vulnerability scanning directly into developer workflows

43:24 for vulnerabilities. So, uh, scan the project for, um, now what might be interesting here is I don't actually have Gripe installed, I don't think. Um, I might, uh, but one of the things my MCP does is it goes to find the binary to see if you've got it.

43:43 Um, and if it can find Gripe in the path, then it will use it. If it can't find it, it will go and get the binary. Um, [snorts] and if Gripe is outdated, it will try and update it. Um, once it's found the binary and it's able to run it, it then has to download a vulnerability database. Every time you run GRE for the first time, it pulls down a vulnerability database, which is like 100meg or something like that. Um, so I suspect it's probably doing that right now. Once it's downloaded the vulnerability database, it will then scan the directory. So now it wants to scan this directory and I say, "Yep, scan the directory." And what it should do is potentially find um maybe some vulnerabilities in whatever dependencies have been added to this Python project.

44:29 And what I would envisionage doing is rather than doing this manually, I would probably put it in my agents.m MD or claude MD to say scan for vulnerabilities before committing any code or scan for vulnerabilities before uploading any code to GitHub because I don't really want to publish it then scan um or you know just test that it works and then scan. I want to make vulnerability detection something that's just part of the process. And by having a vulnerability scanner, whether it's Gripe or any other uh as an MCP server, it means it's right there as part of my developer workflow. It's not an afterthought. It's not something I do later. It's something that I can um integrate right into the process. And of course, and of course there's more you could put into there as well, not just the process, which is super important,

45:17 but also the policies to say like, you know, in the two high, you know, vulnerabilities that have come back here, we don't want any high or critical vulnerabilities. Medium and low is okay maybe for for a specific policy. And you can almost like ask it to not check in anything that has high security vulnerabilities unless the user absolutely says so and so forth.

45:41 Yeah. and and this hasn't given um me a super detailed explanation of what the problem is. I can certainly go and Google that. But if I wanted to ask for more detail, it now has the the names of the vulnerabilities. And so it could describe them to me and give me the detail and say, "Here's what you should change." And that actually can be fed to Claude, my agent, or whichever agent I'm using. And the agent can then fix that. So it says, "I recommend you do this thing here." And I say, "Okay, do it."

46:10 And away it goes. and fixes my code. So, I've found the vulnerability and

Standardizing development with spec-driven registries

46:15 potentially I can have Claude just fix it for me cuz I made a silly mistake and I said, "Use Django 4.0 and what I actually meant was Django 4.2."

46:25 Awesome. So, what's what's next then? Uh with with I think lots of people now using MCPs, um what would you say is kind of uh one of one of the next big things that we should be we should be thinking about? So, I mean, it's predictable, but I'm going to have to say it's spec driven. It's the the the process of creating new software. Like, for someone like me where I can articulate in English what I want, but I don't necessarily know the details of how every single one of the dependencies get built and how they hook together and which APIs I should use. And I feel like if I'm doing that myself and trying to articulate to the agent how to do those things, then I'm kind of wasting time.

47:11 Someone else must have done this already. Someone else must have built this library or used this API. And so I feel like there should be these reusable chunks just like the um the settings and the mcb.json that I found scattered all over GitHub. I feel like there is a real opportunity for those little snippets of um instructions, those specs that we can scatter throughout our project and have that do the work that I would normally have to describe to it or it would have to go and search and figure it out and maybe get wrong. And so I want I want to have something where there is um this repository and that's you know the Tesla MCP and the spec registry is is basically that and that's that's what drew me towards using Tessle is having this repository of knowledge that the agent can call upon to build software in a way that is um secure, efficient, reliable and reproducible.

48:12 Yeah, absolutely. And I think the fact that the the the sharing capability here uh within the within the spec registry I think is I think is very significant because it really starts providing that collaborative approach to to development specifications. Agents are only going to work as well as the specification as well as the guidance is. It's very hard to actually get that guidance that those policies that that style um correct and yes sometimes it's absolutely personal and you might choose to override those certain settings but for a lot of the time when you start getting teams you don't want everyone to to develop in different ways you want to agree on a way agree on a stack agree on a uh a policy that everyone should should abide by and this is where I think this having a having a platform having that that space in which people can contribute, people can pull down and naturally have their agents guided uh by these you know

49:14 policies, workflows, etc. I think it's absolutely core to to to that collaborative development.

6 Highlight · Reusable specifications for consistent agent-driven development

49:21 Yeah. And also it saves a lot of duplication of effort. Like if I uh have I'm cranking out a bunch of little projects, I don't want to have to tell it every time for Python always use UV or always do this. I would way rather have these reusable chunks that I can just pull down and say right here's my stack and it goes and gets the spec for those stack and if those libraries or frameworks have been updated then so should the spec be updated and the spec can be pulled down and if there's a one missing you know I'm a big advocate for sharing and so if there's one missing I should be able to contribute that in a really straightforward way um so [snorts] I can give back because you know I'm a a a uh a proponent of the open source generation and these little snippets of markdown. It's [snorts] it's something that's giving back to those people who are already creating uh these these little snippets themselves. And I I feel like we could all just uh do a tiny bit and build this corpus of best

50:24 practices for uh steering agents because, you know, we we don't want everyone to have to yolo their way through and try and corral and course correct constantly with their agent. You really want to be able to say, "This is what I want. I want to use these frameworks. I want to use these tools. Do it." and not have to keep h well I did ask for that earlier and you didn't do it and I've already told you do this and you didn't do it and so I feel like these specs can keep it on track keep it on Rails and keep the development consistent with the last time you ran it and not a surprise that it's all gone differently and all gone wrong um I I prefer that consistent approach than um than just having an argument with an agent. It feels like a bad use of my day.

51:11 Teslao uh if you wanted to if folks wanted to to to take a look at the spec registry and and and how that can be achieved. So please do please do check that out. Alan, we're we're at time already. Um I really appreciate all the wonderful demos. Um, I think it would be great for people who are who are kind of just getting started into MCP, just getting started into agents to kind of see what's available, uh, learn what others are using and how easy actually it is to really get get started and get enabled with MCPs, uh, and how it really truly does extend um, extend the the the the capabilities of a of an agent. So, thank you very much for showing that.

51:48 Thanks very much. I look forward to uh more of these MCPs appearing and people sharing what they're doing.

51:55 Yeah. And I hope the biggest takeaway isn't that we should all have this nice little audio announcement saying that that that Claude is waiting for human interaction. That's that that shouldn't be something which everyone should do in an office. I can only imagine in an open plan office how frustrating that would get.

52:13 I'll make sure it happens. Don't worry.

52:16 Awesome. Alan, thanks again. Appreciate everyone for uh for joining in. Love to hear what MCP servers you're using uh and if there's anything you want to share and and you know make sure others can can learn from your knowledge and your experience and uh and let us know what you're using. Um thanks again Alan and thanks all for listening and tune in next time.