Sponsored by: Safe Software | All Data, Any AI: Building MCP-Ready Ecosystems with FME

Summary

Safe Software CEO Don Murray explains how FME uses the Model Context Protocol (MCP) to transform complex enterprise data workflows into governed, plug-and-play tools for any AI agent — addressing the reality that 60–80% of enterprise AI project time is spent on data integration.
FME acts as an MCP server that connects AI agents including Databricks Genie to spatial data, legacy systems, air-gapped infrastructure, and real-time feeds that the lakehouse cannot easily reach directly, without writing brittle custom glue code.
The architecture includes governance controls such as column-level data access control and sensitive data handling, letting administrators define exactly which data sources each AI agent can query while deploying existing FME pipelines as governed MCP tools.

Chapters

FAQs

What is MCP and why does it matter for enterprise AI?

MCP, or the Model Context Protocol, is an emerging standard for connecting AI agents to external data sources and tools in a governed, reusable way. It matters for enterprise AI because it lets teams expose complex data pipelines as plug-and-play tools for any AI agent without writing custom integration code that is brittle and hard to scale.

How does FME work as an MCP server with Databricks?

FME acts as an MCP server that exposes data workflows as governed tools that AI agents — including Databricks Genie — can call during a conversation. In the demo shown in this video, Genie uses FME MCP tools to perform a zoning assessment by calculating spatial statistics and writing results directly back to the Databricks Data and AI platform.

What types of data can FME connect AI agents to?

FME can connect AI agents to spatial data, legacy on-premises systems, real-time feeds, and air-gapped infrastructure — data sources that the lakehouse cannot easily reach without custom connectors. This is especially valuable for industries like utilities, transportation, and government where critical operational data lives outside the cloud.

How does this MCP approach handle data governance and sensitive information?

The FME MCP integration includes governance controls that let administrators choose which data sources AI agents are permitted to access and supports column-level data access control to expose only approved fields. This means organizations can benefit from AI agent capabilities while ensuring sensitive data is never surfaced to unauthorized queries.

Full transcript

[00:09] Okay. So, we'll um I'll get started. Um and on today's talk, I'm going to talk about MCP. Um first of all, oh that's not working. Oh, man.
[00:25] First of all, a uh Can you hear me? Yeah. Okay. Yeah. So first of all some statement from our lawyers. Don't worry nothing I'm talking about today is forwardlooking. Everything is shipping. So then anyway I did that. So uh is this
[00:37] working now? Yes. Okay. Um data bricks would like you all to fill out your survey. So if you can do that that would be great. And uh and now on to what I'm going to talk about. So u I'm from safe software.
[00:51] We build a product called FME. and uh I'm going to talk to you about, you know, data MCP and how it fits into your your AI strategy. So, my name is Don
[01:03] Murray. I'm a co-founder and CEO of Safe Software. I've been doing this for 30 years and uh the world of data never gets boring. I can assure you that. So, so one of the things we're all finding
[01:15] is that um organizations are spending more and more on AI than ever before. I don't know about your company, but I know at SAFE, our token usage is going
[01:27] is going like that. Um, and um, and part of the problem is data. 95% of IT leaders say that data integration is where they're spending their time. Gartner puts this number at 13 million,
[01:41] the amount of time people are spending on getting data quality. And um for every big AI project uh 60 to 80% of the the um the time is spent with data. Um
[01:53] and the thing is the model is not the problem. We have better models all the time. Um you know um Mythos Fable you know that we just heard about Gen Genie coming out today and um we have
[02:07] the staff we know what to do. The actual challenge when it comes to AI is is the data. Um, you have, you know, legacy systems. We all have legacy systems that
[02:19] we can't get off of because there's applications that were written 30 years ago that we just don't have the budget to rewrite. Um, we have critical data that's locked inside on premise data.
[02:31] We're seeing more and more companies bringing some of their data back to onrem. they're running critical infrastructure in industries or they're working in air you know air travel um air traffic control the cloud goes down
[02:44] they can't afford to be hacked the term airgapped is coming up more and more um 3 years ago I never heard of the term airgapped and then all of a sudden now I hear airgapped airgapped airgapped and um and then you just have some workflows
[02:58] that are you know too too complex um to use so so and your lakehouse um it's only as good as the data that you can put in it or and this is where MCP is
[03:10] relevant connect to it. Those are the two things that determine how valuable your lakehouse is. You don't want your lakehouse just to be another silo. You want to connect it to all the systems. We all have to work with many systems,
[03:23] many many data um data sources. And this is the data that you know that we've been working with for for a long time. You have your legacy data. I talked about that. You have your spatial data, geospatial, BIM data, where are
[03:38] your customers, where's your stores, CAD and GIS, real time data. We're seeing a lot big uptick in real time data across industries. On premise, of course, the data that's
[03:50] behind a firewall that's never ever going to be allowed to touch a cloud, right? And um and then there's proprietary formats um and then just really sensitive um um data. And when it
[04:04] comes to quick to Yep. Okay, I skipped one. I'm going to go back. Doesn't matter how many times you go over your slides before you present you notice you got one out. And so we recently because we've done geospatial data for 30 years.
[04:16] We recently added geos the complete geospatial data support for for data bricks. And so if you're working with data bricks geospatial data know that you can get um and it doesn't matter whether it's geography
[04:29] um i.e. looks like the globe or geometry and custom coordinate systems um we can do that. So um we've been doing this since 1993 and you can kind of see the
[04:41] evolution of the data we worked with and then in 2025 um Anthropic created MCP probably the biggest thing I've seen in my in my career I remember the first
[04:54] time I got the spec I'm reading through it model context protocol it's kind of like FME doesn't help you when you know what it stands for and then all of a sudden the light went off and I'm like,
[05:06] "Oh my gosh, this is actually an amazing integration protocol. This might be the integration protocol we've been waiting for um my whole career."
[05:18] Um, and it's not, you know, it's not a feature. It's the it's a new integration protocol that can handle all kinds of data. Doesn't matter whether it's unstructured data, structured data, um,
[05:30] spatial data, BIM data. Um, and the other thing is is it's not read only. It's full CRUD. You can create data using MCP. You can read data, you can update data, you can delete data. Um,
[05:42] it's a protocol. It's really a strategic shift. For years, you know, we've all been building these one to point to point, you know, very brittle. Now with M with MCP, you truly do have this ability to control what's on both sides
[05:55] of the MCP, the caller or the creator. and um it's really already uh making a huge impact on um industry. So what do we do as a data integration vendor? People have
[06:07] MCP is the end of data integration." I'm like, "MC is the greatest gift of all time because now it just changed the game because now what we do is we are we just focus on building tools and I'm
[06:19] going to show you in a minute. We have a demo and then any MCP client can call it." So it could be you know the ones that get all the press are is AI but using this you can spatially enable in
[06:31] our case any data set or any application like SAP like um data bricks and and um so it really is um you know exciting if we take a look at before MCP you know um
[06:44] and you're trying to work with AI it was really difficult for the agent to get to perform spatial analysis you would have to build custom code um if the data was behind a firewall, it was unreachable.
[06:57] I'll talk about how MCP now enables you to control what data behind a firewall is reachable. Um, and then now after MCP, you can expose um your data and
[07:10] tools, so processing tools as um to MCP and track who's using them. You can control them with security ooth 2. Um
[07:22] and anything can anything can call it. So I have a demo. There's nothing like a demo. Um and first a slide to explain um what I'm doing here. Okay. Um and I'm not brave enough to do it live. I'm
[07:35] recorded it so don't worry. Okay. Um yeah. So first we have the AI interface layer. Um we're going to show using data bricks genie. Of course it works with claude. Of course it works with chat GPT
[07:48] but this is a data brick show so we might as well show data bricks genie um and then um running in docker uh we run in kubernetes docker or whatever we have um our server called FME flow and it has
[08:02] um you can create as many mcp servers as you want with no code and then for data sources we have some property transactions table and data bricks we have some documents in shareepoint which we're going to um pull information out
[08:15] of to these unstructured documents. We have ArcGIS Enterprise which is a spatial GIS tool. Anybody here know about Argis? Okay. Yeah. Okay. And then we have because we're from Vancouver, we
[08:27] decided let's also pull some data from the open data portal that is sitting around on an FTP site and then we're going to write it back to um to data brick. So that's the demo we're going to do and I'm going to let Dan who built
[08:39] the demo um walk us through In this scenario, we have a real estate development company that's evaluating sites across downtown Vancouver for
[08:52] mixeduse development opportunities. First, we'll ask our data bricks agent to evaluate an address that we're considering for development but require additional information about. This prompt is going to trigger the agent to
[09:05] use MCP tools we created in our FME flow MCP server. The first tool, the zoning assessment tool, is running an FME workspace in the background that geocodes the address provided in the
[09:17] prompt and performs a spatial overlay against zoning data stored in RGis online. Similarly, the view cone assessment tool also runs a spatial overlay, this time
[09:29] against a data set from the city of Vancouver's open data portal. In both cases, FME is not only ingesting data from sources AI can't reach, it's also performing spatial analysis to enrich the data returned to the agent.
[09:43] Next, we'll ask the agent to retrieve permitting documents. Again, this is utilizing an MCP tool hosted in FME flow. FME identifies and reads the contents of the relevant work document
[09:56] and exposes it to the agent, acting as the governance layer between our sensitive data and AI.
[10:17] Finally, the market assessment tool utilizes property transaction data stored in a datab bricks delta lake table. FME takes the geocoded address again and searches for property transactions within a 1 km radius.
[10:29] FME also calculates relevant statistics such as the median price per square meter, transaction volume, and price trends on the fly.
[10:52] Now that we have all the information we need, we can ask our agent to submit our results to the site pipeline table in datab bricks for further consideration. Our agent will ask us to confirm the submission before we load it into the table.
[11:15] While the agent is running, let's take a look at our MCP server we created in FME flow.
[11:30] FME makes it easy to create and host MCP tools, all without the need for custom coding. All created in just under an hour. Back in data bricks, we can see the results of the site pipeline update tool
[11:44] and we can verify the results by taking a look at our table in data bricks.
[11:57] In this example, FME connected systems that would otherwise be completely isolated from AI and allowed us to enrich our data and decisions with spatial and statistical analysis, all without the need for custom coding.
[12:15] Okay, there you go. So, that just shows, you know, the power of MCP being able to pull multiple things together from different systems. Um, and really the Brett is the point. Now, with MCP, you have one interface to connect to many
[12:28] different systems. In the past, you would have to figure out how to talk to all those systems in independently. Now it's very easy to do that and you can focus on um leveraging AI or leveraging other tools knowing that all the other
[12:41] data is being um handled by um in this case FME with MCP. And the other thing is you get to choose the AI. There's nothing about um this demo that you could you know swap out AI. Let's face
[12:54] it the industry is moving super fast. um the the the models are becoming a commodity and it's it's the context that matters, right? That's the thing that has the value and how do you build
[13:06] context? You build context with data and so yeah, so that's really um that's it. Um so this is the you know you use this this um with MCP the AI infrastructure
[13:19] decision that really enables you to just build on top of it. It build it creates a foundation that then you can continue to build on top. In today's world, you you cannot afford to have lock in. You
[13:31] need to be able to decide the tools you want. And with MCP, you could even, you know, one system might change, but if you keep the MCP tool the same, um all your workflows you built on top of it,
[13:44] um are not going to uh to change. So, so the whole thing is about choice. Um in today's world um with with our technology you can deploy anywhere you want on premise in any cloud kubernetes
[13:58] it doesn't matter you can pick whatever AI model you want what is the best AI model it depends it depends on what you're doing and depends on the data you're feeding into it um and then of
[14:10] course any data source we do thousands and so you have these legacy systems the data that's hard to reach and that's what we bring to MCP so now you can build a modern stack sitting on top of
[14:22] um MCP. Um last but not least of course is the governance. You need to know who's accessing your data and with an MCP server you know exactly who is accessing what piece of data. Um and the
[14:35] sensitive data never leaves um your control. So one of the things we we make the point of with FME is FME is the data guardian. You might have some data on prem that you're not allowed to share,
[14:47] but you, you know, you could have a table with sensitive data, maybe four columns you can't share, but but the other six you can. You can fully control the data that that you share. Um, MCP
[14:59] just isn't for the cloud. We have organizations who are running MCP inside their firewall with AI that's inside their firewall, with data that's inside the firewall. So no data ever leaves the
[15:12] firewall and so that's the um you know the fully onrem one and so you get full control over who has access to um all your data and the point is the raw data
[15:24] is never ex um accessed directly or exposed. It's always going through that MCP interface that lock that logs everything and the MCP tool you built
[15:36] defined precisely what is going to be shared. So there's no way it can get access to other columns or data because the tool itself never had access to that data.
[15:52] And um yeah, so it's all about data access. And with MCP, the other thing to remember is it's full CRUD. So you can create data, read data, update data, and delete data. If you think MCP is just about data reading, um you're missing
[16:06] the you're missing you don't want to forget that. Okay. So yeah. So there we go. And um yeah, so we're at booth 432. Um come on by and we can talk more. Um share, you know, um Dan is actually
[16:20] there. Um if you want to explore FME, it's fme.safe.com. And uh with that, um thank you. Yeah. Yeah.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.