Transcript#

This transcript was generated automatically and may contain errors.

All right, we are one minute past the hour. Thank you, everyone. Really appreciate your time today and excited to talk to you about MCP in practice. So I'm Matt Leary. I'm a product manager here at Posit and I work on Posit Connect. And I'm joined today by my colleague Samuel. Samuel, do you want to introduce yourself as well? Yes, sure. Hi, my name is Samuel Hawk. I'm a solution architect in the HLS enterprise segment at Posit. I support our customers building AI workflows. Great. Thanks, Samuel. Well, together, we're excited to host this workflow that's going to focus on or this workflow webinar that's going to focus on using MCPs in your workflow.

So before I jump into a couple of slides, just lay out what you can expect today. I'm going to start us off and cover some basics of MCP and also review how you can host MCP servers on Posit Connect. And then I'm going to hand it off to Samuel. He's going to then do a walkthrough of workflows using MCPs and show you some code and different things with that. And then our intent is to leave about 15 minutes at the end for Q&A and we'll wrap it up there. And for me, I'm really excited about the Q&A, both to hear your questions, but also hear the use cases you're seeing, what you're doing with MCP, and just have that discussion at the end as well.

What is MCP?

All right. So let's just start off with some level setting of what is MCP. So MCP stands for Model Context Protocol. It was developed by Anthropic in late 2024, and it's now under governance under the Linux Foundation. So I'm going to walk through a basic example in the next slide. But for now, I think it's key just to highlight that MCP is a widely adopted and open source protocol, kind of similar to REST APIs. It's a specification that defines how things can work and how your AI clients can consume tools. And there's a couple of key terms I just want to hit right here before we move on. The first, as I just said, MCP is a protocol. It's the standards itself. It defines how a client and a server can talk to each other. Moving over to the second box there, MCP servers provide tools. That's what you host to write tools, prompts, and resources to an MCP client. Now, I should note the specification, and we'll share that in the link so that if you want to look at it, you can see all the details on that. It talks about tools, prompts, and resources, but largely what I've seen and people are using are the tools.

And then lastly, you have MCP clients, and that's what consumes the tools from the MCP servers. Now, they're called MCP clients, but they're also often called AI clients or AI applications. But you can think about that as Cloud Desktop, Chat GPT, Posit Assistant. It's an AI front end that is able to access and use those tools that we mentioned hosted by the MCP servers.

And then lastly, I want to highlight a couple other distinctions too. You can host MCP servers either locally or remote. Local is exactly what it sounds like. It's running on your own computer. And what's nice about that is you don't really have to worry about the security as much, but you also can't share it easily with colleagues. Remote servers means the MCP server is hosted outside of your computer and you're connecting to it. That means you really have to consider security and authentication, but it also makes it easier to share. We're going to focus on remote MCPs for this presentation. And as part of that, we'll talk about how Posit Connect makes it easy for you to host MCP servers. And the last part, before I kind of go over an example of MCP, server MCP versus skills. That's probably something if you're looking at MCP servers you're talking about. We're going to focus on MCP for this webinar, but there's also a lot of times where a skill or a CLI might be the more appropriate way to go to solve your business problem. Like I said, we won't go into it, but it's important to note that MCP servers can solve a task, but they can also be used with skills, with CLI tools, and you can pair them effectively together.

MCP in action: a classifier example

All right. As I mentioned, I want to quickly go over an example of like what's behind an MCP server. So let's start on the left-hand side where we have our MCP client. In this example, it's Posit Assistant running in Positron using one of the cloud models. So that's your MCP client. That's where you're doing work. In this example, we built an MCP server that has one tool that's a classifier tool. So what does that mean? It means that we have a tool that's going to call out to a large language model on the right-hand side, and we've had a customized prompt and some instruction so that it can classify text in a very specific way. So your MCP client can connect to the MCP server, see there's a classifier tool available, know what inputs it has to pass along for it to be able to operate, and then at the end of the day, the MCP server will go out to the external system, which is an LLM here, get the response, and return it to the MCP client.

Now I'm doing a pretty basic overview, but I think what's really powerful about that is for the individual who's using Posit Assistant, you don't need to worry about the back-end. What is the tool doing? Where is it accessing this LLM? What is the prompt? Did somebody update the system prompt last night? All of that is handled for you. You just know you have access to this tool, and your AI client can call it, and it knows what to do with it and how to get the instructions back.

All of that is handled for you. You just know you have access to this tool, and your AI client can call it, and it knows what to do with it and how to get the instructions back.

When to use MCP servers

We did want to cover what you should consider before you build an MCP, and this isn't an exhaustive list by any means, but we'll kind of go through some use cases where MCP works well, and maybe where you want to be kind of careful.

So where it works well, you want to share tools across groups and different AI clients. As I mentioned, MCP is a standard protocol, so if you have business users who aren't writing code in an IDE's operating cloud desktop, they can connect to the same MCP server that you connect to with Posit Assistant. So that can be really powerful that it extends where the MCP server can be consumed by individuals. It also makes it easy to distribute and update. You could share skills across a cloud desktop and Posit Assistant, but you have to keep those up to date, and again, with a remote hosted MCP server, you update it, and your users don't even have to worry about it. They automatically have access to the new functions.

Another use case where it might be helpful to use MCP is if you have tools that have specific resources or packages. I'll go back to it. If you have a Python-based tool that's calling NumPy or Pandas under the hood, if you have a user using cloud desktop, how do you know if they have those packages installed and they have proper access to it? You can put these different requirements and resources on a remote MCP server, and again, your users don't have to worry about that. That's all hosted centrally, and you can manage that.

You have credentials you don't want to share. I can use an example that we use here at Posit. We use Zendesk as our help desk, and you can access that via an API key. Well, we might want multiple developers to have access to that to look into different issues and look at code at the same time. We don't want to have to distribute API keys to everyone and then manage and rotate them. We can put the API key behind an MCP server and then monitor the use and give access via the MCP server, and nobody even has to be aware that an API key is being used.

And then lastly, it shines where your tool returns small targeted results. I actually talk about that where to be careful about why that's so important. But as an example, let's say your organization has datasets on hundreds of clinical studies, and you just want to know, do we have a dataset that has this type of information in it? Well, if you have somewhere stored the metadata around the different datasets you have, that would be a good use case for an MCP to go out, search, and say, yeah, you have access to these three different datasets, and then you could independently go out and look at that data.

All right, so that's kind of where it works well. Again, not an exhaustive list, but some good examples. Where to be really careful. We need to call out here, you are using an AI client, so a large language model on the front end. So every tool call and every tool that you bring into your context, it's going to eat up tokens in your context window. So you're losing space in your context window, and you're also paying for that. I mentioned before about small target results. MCP is not great at moving large quantities of data. That's because the data is going to pass through tokens, and again, there's a cost to that, and it's really not an efficient way of doing that.

And then lastly, a skill is often cheaper and quicker depending on what you're trying to do if you're teaching or instructing an AI client how to do something. I'll use the example of coding standards. If your organization uses CamelCase and you want to document functions in a certain way, you could pass 500 lines of code to an MCP server, but that's highly inefficient and slow. A skill would probably be more efficient there, so your AI client knows how to generate the code that meets your company's standards. And then lastly, I mentioned about, I think it's worth reinforcing, remote servers really allow you to use shared computer resources. They're easier to update, and your end users don't have to worry about it. And then more importantly, they give you centralized logging and insight into how they're being used, which if they're all running locally or if people are downloading skills on their own computer, you don't really have that insight of how it's being used.

Hosting MCP servers on Posit Connect

Now I want to move on to how we've been making Connect a great place to host remote MCP servers. And here's some of the things we've been doing for that. First, auto-detecting MCP servers based on popular frameworks and listing them out on their own tab so they're easy to find and discover. And we also set a minimum process so that if someone's connected to the MCP server, it's warmed up and you don't have a 10-second lag as it's spinning up, because that can be strange for the user who doesn't understand why it's taking so long to respond.

We also now have an interactive landing page. If you've built an MCP with FastMCP, which is a very popular Python framework, we auto-generate an interactive landing page with info on connecting to the server and what tools are available. We'll show you this later, but it's possible because, as I mentioned, MCP is a well-defined standard, so we're able to inspect the different tools and pull that information out. I'm really excited about this feature, but it can be really helpful to quickly see information that was hard to get at before if you didn't build your own custom landing page, which you can still do, but we just want to do that out of the box for you as well.

Additionally, Connect comes with a built-in OAuth 2.1 server. This allows you to authenticate to an MCP server hosted on Connect via browser-based flow. I'm going to show that on the next slide. This means no API keys that you have to manage and rotate. It's something your security teams will appreciate, and it's the recommended flow that we, or the flow that we recommend. We also have built-in usage monitoring. You can quickly see how many people are using my MCP and how often. We think that's really valuable. You can see the value that people are getting out of the MCP. And speaking of access, you can manage access to your MCP via access control lists and groups already in Connect. You can integrate Connect with your identity provider, such as IntraID or Okta or many others, making it easy to manage that access with the groups your organization is already using. And then lastly, we spend a lot of time building out integrations with popular third-party services, such as AWS, Azure, Databricks, Snowflake, and you can use these to connect those resources securely.

I do want to note, a lot of these things, the OAuth server, the ACLs, the usage dashboard, they're not just for MCP servers. They're built into Connect, and you can use with other content types. So that's some of the ways we're trying to make Connect a great place to host MCP servers. For my last slide, returning over to Samuel, I wanted to show you how you can connect to an MCP server on Connect by that browser-based flow.

Connecting to an MCP server via browser-based OAuth

All right. Let me get this started. Okay. You can see here in this example, I'm using Posit Assistant and Positron. I have this repo, and I'm going to ask, what MCP servers are configured in this project? Now, I know the answer. None are configured right now, so it's going to come back and tell me I don't have anything configured. And so a very standardized way to configure and connect MCP servers is with a settings.json file. So you can see here, I'm going to my Posit Assistant settings.json. I'm pasting in about nine lines of code that says, hey, here's the URL. Here's I want to connect to it. Here's the browser-based flow where I'm just going in, authenticating to the Connect and to the MCP server through OAuth. So what this means, again, no API key that's hard-coded that I have to manage and have to keep secure. So again, that's the recommended flow that we suggest for connecting MCP servers. Again, your security team is probably going to prefer that method as well. You can see here now I asked the same question about what MCPs are available. It's listing out all the tools, and now I can connect to the MCP server. So it's something that's been very popular and I think is a really nice way to connect to MCP servers hosted on Connect.

All right. With that, I'm going to stop sharing. Thank you for that time. Just kind of do an overview. Happy to answer more questions at the end, and Samuel, I'll turn it over to you.

Building an MCP server

Thank you, Matt. Okay. So Matt gave a great overview of how we can deploy and use MCPs. So let's see how it's been done, right? And we'll start with how do we even build an MCP server in the first place, right? So if you're familiar with building APIs, it's pretty straightforward because, as I mentioned, FastMCP provides a really nice wrapper. So you're kind of building it very similar to FastAPI tool calling, and you can deploy it. But where I love to get started is this documentation of MCP servers within PositDocs, right? Because it walks you through in various aspects. For example, if you already have a local MCP server, and how do you want to convert it to a remote MCP server and serve it across your organization? It has very much detailed documentation, and you'll find really get-started examples, right? And basically, for most of your tool calling, either you can add a decorator, or there are other options to add custom tools by definitions, and then just define your Python functions as a tool. That's pretty much it.

What I do then is I go to my Posit Workbench, right? So for those who are not familiar with the Posit ecosystem, Connect is where we host our applications, APIs, and MCP servers. And Workbench is where we do most of our development. So if I open a Posit transaction on my Workbench, I can literally go to Posit Assistant and provide the description of the MCP server I want to build, and I just ask it to get started. So this flow has become very popular for the users who are building MCP within Positron. So this is where you might want to get started. So literally, just come here and give it a try. I think that's the best way to get started with the MCP server.

Now we're going to go over a few examples, but before I go there, I wanted to show you an MCP server that I have built over here, which talks to Snowflake. And the idea is sometimes,

if you're in life sciences, either your clinical data or your EHR data, they are landed in Snowflake and they might be updated very frequently, right? And you may want to serve some key insights from the data source. It could be a new adverse event or a total serious adverse event you're observing. And you want your AI tool to be able to grab those results, right? And you don't want the AI tool to go and execute a random SQL query, right? So the idea is, if you already know there are some structured SQL query or functions that you trust or that might have been validated by your data scientist, you can just convert them into MCP tools, right? And deploy the MCP server into Connect.

I wanted to highlight that for integrating role-based access control for Snowflake, right? That's where OAuth on the Connect comes into play. Through Posit SDK, we provide some great tools to ensure that authentication is seamless, right? And we'll share the documentation to this SDK later in the chat.

Deploying an MCP server to Connect

Okay, so if I build an MCP server, then how do I deploy it? It's really a single click. So here is the deploy option in Positron, which is Posit Publisher, and you just deploy your project. You select which Connect deployment you want to use and deploy it over there. So let's see how things look on the Connect site.

So I am on my Connect deployment, and you see I have many different deployments here. Say I just want to see what are the MCP servers that we have here, and then it will just slice and dice my MCP servers that I have deployed. I can also see, because I have admin privilege and some of the MCPs are shared with me, what other teammates are deploying in this Connect instance.

What I really love about deploying this Connect, before this feature was available, I had to do a lot of debugging, right? I deploy it and then I test if I'm getting back the right result. What I love about this is the landing page, right? So now I have opened an adverse event MCP server, and this is the first example that Matt was talking about, right? So if you have a free text adverse event narrative, we created a customized SLM harness as a tool call, which basically structured it into a MedDRA definition. So for those who are not familiar with MedDRA, it's a controlled terminology to share different adverse event narratives across the healthcare and life sciences ecosystem.

So here you can see I am getting a nice view of what tools are available here, right, in the adverse event summary. Actually, I opened the wrong one. So this is the pharmacovigilance, where we have the classified adverse event tool. So here you can see I have three tool calls here, and I can ask. I have copy-pasted a sample narrative, but this is a patient narrative that someone has, right? So I can run this, and what it will do, if the API token or the credentials that I am using here is valid, it gives me this nice classification because that's how we have constructed the tool. So imagine using this from any client, right? It doesn't need to be positive system. It could be your cloud desktop or other custom AI client you're building.

So what's happening here under the hood, it's calling another model in MedDRA, right? Now you may ask, okay, how do we connect to MedDRA from this deployed MCP server, right? Again, there are two different ways. In this case, we consider a scenario where you may have a service account, right? So you want to use that same service account, and what you'll do is you can embed your AWS XSK token, so that's one way. But there could be a better way, right, where you may need role-based access control.

That's where we move into a second example, where we look into really getting some adverse event summary and adverse event details. In this case, we are actually connecting to a Snowflake instance, right? So again, I think I have already shown this MCP server has six tool calls, right? A overview, a summary. If you want a description of a table, it can describe those tables. You can get a preview of the table, as Matt was mentioning. We don't want to pull a really large table here, but if you want to see a sneak peek, just the first five rows or things like those, you can see those, right? And then if you're interested only in the serious AI listing, right, you can get them.

Now, how does this integration work, right? So on the Pranav MCP, you saw I used token, but here you see I'm using a native auth integration that's available to me. And I can actually have many different types of custom integration, including Snowflake and Databricks, right? So you can imagine a scenario, you may have disparate data sources, right? And you may want your, or disparate tools, and you may want your MCP servers to connect to multiple tools. That is possible, right? So you can really use your use case and imagination to get creative. But once I log in or connect through OAuth to this MCP server, it will also trigger the Snowflake authentication as well. Because in that case, we'll reach out to Snowflake as my user role, not as a service account. So the role-based access control will be maintained.

So the role-based access control will be maintained.

So if I ask that I want to see an area overview from an ongoing study, so in this case, my ongoing study is CitiesPilot1, because that's the one I have access to. But basically, what it's going to do is, I think I have mistaken something. Okay, yeah, there was a typo, so the study was not found. So in this case, it's really going to Snowflake and bringing back the results live to this, right? So now, just for testing, I deployed it. I have not yet shared it with my entire organization, but here I can really nicely test what's happening.

Matt, you want to add something? If you don't mind saying, well, just because I'm such a big fan, this is a new feature in Connect about two months ago. Exactly what you're saying. We add this in because what we were hearing is, you know, you deploy an MCP server, then you go back to your MCP client to connect, and debugging could be a little bit difficult. And so we put a lot of really nice information here where you can test it and get the response you can see there. The connection strings at the top just make it easier so that you don't have to remember, like, well, I'm using Cloud Code here. What's the format? And now I'm using Positron. It just makes it very easy to connect to it. And I should note, if you want to make your own custom landing page, you can. This will only be displayed automatically if you don't already have one. So you do have the ability to control that, but it's really nice that just out of the box, it will get this information. I know I find it very useful, like you're saying, Samuel, to test it out and make sure things are working.

Yeah. Actually, on that note, that's what I was going to show as well, because this is the default landing page, right? But Matt also created a few weeks ago a custom landing page for this use case we are working on, which is, again, CDISK SDTM validator, right? In your case, probably you are reaching out to some standardized API call that your organization already uses for validating your dataset, right? But you can think now, you can also provide these tools to AI clients through these MCP servers, right? So if you create your custom CDISK SDTM validator, you can also create a custom landing page. And this is what you are seeing here. So this page is slightly different than what I was showing before, because this is a custom landing page, right? So you have both options available. It's very flexible.

Using MCP servers from Posit Assistant

Okay, so now we have the MCP server. How do we use it? What I'm going to show first is how can we connect from Posit Assistant, from Positron, and how does the experience look like?

So here I am in my Posit Assistant, and I'll show you my settings.json, because within this project, I'm connected to these three different MCP servers, right? So I have this pharmacovigilance MCP that I covered. I have CDISK MCP, and I have adverse event MCP. And you'll see, like this CDISK MCP sits on a different instance of Connect. So it doesn't need to be all one same Connect server. You can have multiple Connect instances within your organization and connect to those.

If I ask, okay, tell me about adverse events in the CDISK pilot one study. Okay, maybe I want to be more specific. I want to know the number of adverse events. So now you see it's immediately triggered, that MCP tool call, and it's asking me whether I would allow this call. And what I'm going to do, I can allow it for once, but I'm going to allow this MCP adverse event server for this session. Okay, so now what's happening is basically it's going to Snowflake and running some live query on behalf of me, and then it's getting back the result.

There you go, right? So we have now a summary of the adverse event that has been reported, right? Just to confirm that, who am I, right? I can also ask, who am I? Now it's again asking to call another tool, which is called, who am I? And I'm going to allow this for this project this time. Okay, and it's reporting. Snowflake sees me as samuelhawk at posit.com, right? So this is my user account. So, and it's assuming the role that I'm supposed to, that's my default role, and it's using the warehouse that I asked to, right? So you can have really granular level access control on this tool calling, and that's the beauty, right? That's why we want for enterprise use cases to use this tool.

Similarly, another thing that I love, Posit Assistant is context-aware, right? So here I have an adverse event narrative dataset, right? So maybe you are just working with this data, and you want to, maybe you load this. Okay, so I want to source this R file, and I want to see my variables. Okay, so this AE data is loaded, and you'll see this dataset has some AE variable team, and I'm basically looking at different subject ID and their report of their different adverse events, right? So I can come on, and let's see. I can create a new chat, and can you, so I'm asking, classify the adverse event for subject ID, so maybe I'm interested in this particular subject, okay? And I just want to see what the major classification would look like. So what it's doing first is trying to get the data, get the data, and it's filtering the adverse event, and then you see, after it filtered out the data for that particular subject ID, now it's going to call this pharmacovigilance MCP, right? So I'm going to allow this call, and then what it's going to provide is the MedDRA classification that we already seen on the Connect side, right, by testing the API tool calls, and you can see this is the classifier output, right? So this is what exactly came out of the MCP tool. For some reason, Opus also thought that it needs to provide its own read to that, which is fine, I don't mind it, but this is the classifier, actual classifier output, and this is what Opus kind of, its own explanation of the model we are using.

So you can have really granular level access control on this tool calling, and that's the beauty, right? That's why we want for enterprise use cases to use this tool.

Connecting from Claude Desktop

Okay, so you might be saying this is great, but not all of my business users are going to use Posit Assistant, right, because you may have your enterprise subscription to certain desktop client, right? How do we use it? So you may have noticed, let me minimize it, in this page, we provide instruction to how to connect to different clients, right, including cloud desktop, cloud code, or any other custom client that you could think of, right? So what I'm going to show next is how does this work on cloud desktop.

Okay, so I have a small recording that shows how do you basically copy this URL, and of your MCP server, right, so you're here, you can create a vanity URL, copy this URL, so this is the CDisk validator MCP that we discussed, and then on your cloud desktop, basically, you add it as a custom connector, right? To do that, you give it a name, because it's a custom connector, and then you just provide this MCP server URL, right? And depending on the type of authentication you are going to use, there will be different options. In this example, we did not require a sign-in, it's a more flexible MCP tool anyone can use, and then you can also see the tool listing, right, on your cloud desktop.

Now you have this connector, right, and you can just start asking questions. For example, if you have some CDisk data or metadata, you can drag it here on your cloud desktop, and ask it to validate the dataset for you. So now what it's going to do, again, a very similar pattern, right? You allow cloud to reach out to this custom MCP tool, and then it's going to

search for the available tools, see what's there, and bring back the result on cloud desktop, right? So again, this kind of showcases how you start distributing the custom tools that you build to enterprise-wide AI clients by hosting it and connect.

That covers the key component that we wanted to cover today. Matt, back to you if you want to add any additional thoughts.

No, thanks for that, Samuel. I think just, I like how you showed the cloud desktop version right there too, because I think that really highlights one of the key reasons why MCP Servers can be really helpful for you. Because again, it can connect to different cloud desktop, Positron, you can meet users who are using different tools in the standardized

Monitoring and usage tracking

Yeah, I actually wanted to cover one more thing that I forgot, because using itself is not enough. You also want to know who are using it and how.

Right, so if I go back to my, let's see, there you go. So if you click on the settings, right, you have these monitoring options. So you can actually monitor how your total visit, who are the visitors. In this case, I am the only visitor, so one unique visitors. So you can kind of have this monitoring capabilities to see what's happening to your MCP Server. As an admin, you can go to the admin dashboard as well. So if you are connect admin, you can actually have a broader look of all the MCPs that are deployed and who are using it. You can drill down and see users pattern of each one of these different MCP Servers.

That's great timing, because we actually had a question about tracking user calls. So that's exactly what Samuel just showed. We can see how often your question is about seeing the individual tools and what they're doing. That's not built into that usage monitor, but it's something that we're looking at and trying to figure out how to make it even easier for everyone to have the insights into that type of granularity.

Q&A

Now we will start the Q&A at this part. So we'd like to thank everyone who's been sending in questions. And I'm going to go down the list of some we have. And the first question I see here is, is the R package mcptools recommended for deploying fast MCP to connect on Posit Connect with R? So one little nuance there. So fast MCP is a Python framework that you can absolutely deploy to Posit Connect. mcptools is a Posit package that we've built that is for R. And yes, you can deploy R-based MCPs to connect with the mcptools package. It also has some other components that you can do other things around MCPs, like make it easier to build an MCP client as well. So fast MCP, a Python framework that's really, really popular and used heavily for MCP servers. And mcptools is an R-based package that can do different things around MCP, including deploying an MCP to Posit Connect. Yeah. And just want to add, it is supported, right? And it uses a tool calling based on LMR. So if you're using LMR already and have LMR tools built in, you can bring them quite easily to mcptools ecosystem.

All right. Our next question, Samuel, I'll take it at first, but I'd love to get your take on it too from a practitioner standpoint. But what's the minimum memory footprint for an MCP? Since it's recommended for it to be always on, I'm curious what the overhead will be.

So what I'd say there is we auto-detect it and turn it on to a minimum process of one. But there might be a use case where you can adjust that. You could turn it down to zero. Maybe you're just testing it out and you know I use this MCP once a week. You probably don't need to set that to one. We also have some intelligent logic built into Posit Connect that at a certain point, it won't keep setting every MCP to one because eventually you'll max out processes like that. So I'll start by saying that in most cases, it probably makes sense to have them turned on, but that's something you can configure based on your use case. And then same thing, I would say it depends in terms of how much resources to put behind it.

If you have a tool that just a few people on our team are using, you probably don't need that much resource behind it. But it also depends on what it's doing behind the scenes. If it needs more memory because it's doing some large computation. So it depends. But what I would say is the nice thing with the usage dashboard Samuel just showed, take a look at that. You can see are 10 people using it? Are 50 people using it? Are they using it 18,000 times a month like one of our internal MCP servers? Or is it 50 people using it twice a month? So I don't have a direct answer in terms of like, oh, here's what you should set. It's really going to be based on your use case and then probably monitoring and adjusting as you go along. Samuel, would you agree with that or anything to add there? Exactly. That's what I'm going to tell. It depends on the use case, right? If you have a very popular MCP server that's used every day, probably you need to have it turned on and scale it accordingly.

The next question, and Samuel, I'll look for your help on this as well, but what environment variables do you need to set if the Connect instance is authenticated to Bedrock via an IAM role policy? I'll let you, do you mind tackling that one? Yeah, sure. In my example, I used a temporary live access token that on my Bedrock side, I was setting a time limit. You can also get a service account, right? One thing that I want to highlight, and this has nothing to do specifically with MCP, but with Connect. Connect supports adding IAM integration, like so you can bring in your IAM, like as you saw my Snowflake integration, you can also have your AWS integration. And as an IAM role, or you can have also your AWS service account. So it could be a very similar flow, or it could be API token based. So both patterns are supported.

Great, thanks. I'm going to read out this next question, but I'll respond directly to the person probably outside of this, because it might be a nuanced discussion, but in case everyone else is having this, the question is, does the OAuth authentication flow between Posit Connect hosted MCPs and Positron and Posit Assistant currently only work on Positron desktop? And their setup is using Workbench, and they're saying they're having some issue doing the authorization.

So that is something where you have to set a redirect variable in your instances to make sure that flow goes correctly. And we have some documentation on it, and there was a version of Workbench that had some issues. So I'll reach out to that individual and ask a couple questions so we can answer that. But yes, that might be a configuration setting in your Workbench and Connect instance where you have to tell it where the redirect goes. And we do have documentation on that.

We had a question around, and this is an interesting example, I'll maybe let you go first, or happy to take a crack at it if you want, about the use of custom MCP servers versus third-party MCP servers. Any thoughts around that? Yeah, I mean, again, I think in pharma, right, probably custom MCP servers, think then about custom tools or validated statistics or queries that you trust, right? Because you build it, you can vet it, you can get other people on board very easily, right? So when you use third-party, right, either you have to trust completely whoever provided that to you, or if you're using MCP provided by another vendor, you may already have an agreement with that vendor, and you trust them enough so that you use their tool in a regulated environment.

And I'll be frank, I think for most of the custom workflow, right, the way the live services industry has evolved, you probably have your own custom SAS macros, right, your own R libraries that you don't expose to the world. Rather, those lives as your internal GitHub repo and maybe connected to the opposite package manager. Think of MCP tools like that way, I would say, the custom ones, but if there is something already available that you can vet or you can trust comes from third-party, definitely go use them, right? Like at Posit, we trust some connectors, right, like I can connect to my Gmail, right, like because we have that trust, we have that agreement with Google, so we allow, and also with Anthropic, right, so we allow Cloud to read my Gmail and summarize my email, right, so it again depends on your organization's security and IT policy and overall how you look at AI governance.

Absolutely, agree, and one other thing I'd like to note too is a great example is GitHub has an MCP and to Samuel's point, it often seems better to have targeted, you build your own custom one because the GitHub MCP might have a hundred calls. It was actually well known early on that it had a bunch of different tools that you could invoke, so that was again eating up a lot of your memory context and maybe you don't, maybe you want your AI client to have the ability to read issues but not write issues, so in that case you can take a third-party one and then Samuel kind of mentioned you can adjust it, you can edit down, you can fork that copy, have your AI client say, hey, remove all this, we only want these couple of skills, so that's something else that I've seen as well as taking an open source third-party one and just taking out the components you need because they're building an MCP for a bunch of different use cases and maybe in your organization we only need these seven tool calls, we don't need all these other ones, so you can take open source third-party ones and then change them to fit your needs and add new tools or remove things so that it really customized for your organization.

Like at Posit, we trust some connectors, right, like I can connect to my Gmail, right, like because we have that trust, we have that agreement with Google, so we allow, and also with Anthropic, right, so we allow Cloud to read my Gmail and summarize my email, right, so it again depends on your organization's security and IT policy and overall how you look at AI governance.

I don't know much about this but I found some documentation, but there's a question that came in Samuel, I'll ask you if you know, about integrating MCP servers with Microsoft Copod, I'm assuming that means Copod 365, do you know anything about that? I have not explored that, that's a good question. That is, I will put into the channel because I looked it up as I saw that, there's some documentation on the Microsoft site about Copod extending it with MCP, but we'll admit I just found the documentation, looked at it briefly, in terms of implementation your security might have standard or you'd have to really talk to probably to your security team as well, but there is some documentation on Microsoft to share, I'll post that there.

Great, well we're getting close to time, I think we have time for at least one more question and this is one Samuel, again I'll maybe ask you if you want to take the first cut at it, but there's a question around to Posit Assistant and being used with PHI, can it be used with PHI? Yeah, so Posit Assistant when used from our professional product which is Posit Workbench, right, or also with our free version desktop, Posit Assistant itself does not store the data, right, we do not, but it's also important to understand when you are sending the data to an LLM, right, that LLM can see parts of the data, right, whatever, or the metadata, whatever you are sending. So in an enterprise setting what is expected, right, ideally you have a zero data retention policy and agreement with your LLM service provider, that's how I would say majority of the licenses organizations operate, right, so they probably have some kind of agreement with this LLM service provider, where we do not, so they do not retain any data.

Now Posit also provides something called Posit AI Pass, right, and Posit AI Pass also comes with this zero data retention policy and that's a subscription, right, so you can also use that and that will not store your data, but LLM can see your data, right, that I think is important.

I think one thing to add there to the Posit AI Pass is really for consumers, that's not really an organizational thing, our expectation, especially here for enterprise accounts or organization, is it is a bring your own provider, bring your own key, as Sam was saying, so that's really up to you, what you feel comfortable in the agreements you have in place with your provider versus with Posit Assistant directly.

I guess actually we have time for one more question. There is a question about the languages that are supported for MCPs on Connect, so we mentioned R, we mentioned Python, we do on the advanced tier also have Node.js, so that is an option as well, that's a recent addition where you could build a Node.js MCP server and host that on Connect as well.

Great, well we are one minute out, so I think this is the perfect time to wrap up. Again, thank you to everyone who took the time to join us today. We have a couple of questions we'll respond to outside after this, especially as it related to that workbench and getting connected to an MCP server on Connect, so we can get you unblocked there. But from my side, I really appreciate you taking the time to spend with us today. Feel free to reach out with any questions and yeah, have a great rest of your Wednesday. Thank you everyone. Thank you.