Implications of Using HKDF as a Key Combiner

The big shift toward universal ai connectors

Ever feel like you are drowning in a sea of custom api integrations just to get one ai tool to talk to a database? It's honestly a nightmare.

But things are changing fast because of the Model Context Protocol (mcp). Think of it like a usb-c port but for ai models. Instead of building a unique bridge for every single data source, everyone is starting to use this one universal plug.

• it's the new standard: anthropic released this in November 2024, and now even openai is jumping on board. (20VC x SaaStr: The AI Power Grab) it's becoming the way agents actually "see" your data.
• killing the integration tax: a 2025 report by TribalScale shows that traditional setups take months, while mcp can cut that down to just three weeks. This "tax" is basically the wasted time and money spent on custom code that mcp finally eliminates.
• massive ecosystem: there's already over 5,000 mcp servers out there for things like github, slack, and postgres. (Top 12 MCP Servers for AI Models in 2025 - LinkedIn)

I've seen teams at places like Salesforce and Symmetry Systems move toward this because it stops vendor lock-in. If your security agent can talk to any mcp-enabled tool, you aren't stuck with just one provider anymore.

Anyway, this shift is huge for how we build infrastructure. Next, we'll look at the specific saas giants leading the charge.

SaaS giants and crm platforms leading the pack

So, why is everyone from salesforce to hubspot suddenly obsessed with mcp? Honestly, it’s because they’re tired of their customers complaining about how hard it is to get ai to actually "see" their data without a massive headache.

I saw a post by Riccardo Tartaglia where he explains that mcp is basically like giving your ai "hands" instead of just a voice. For these big saas companies, it means they don't have to build 500 different connectors for 500 different tools anymore.

It isn't just about being cool; it is about survival in a world where agents do the work.

• standardized context: companies like notion and confluence are making mcp servers so an ai can search your docs without a custom api mess.
• speed to market: as we saw with that TribalScale data, we're seeing those integration timelines drop significantly.
• ecosystem play: Jason Andersen mentions in forbes that saas providers benefit because becoming an mcp server lets third-party agents work with their datasets way easier.

I've noticed that even google and microsoft are turning their apps into mcp clients. It’s making that old integration tax feel like a thing of the past. Next, we'll dive into the developer tools making this happen on your laptop.

Developer tools and ide ecosystem

Honestly, if you're a developer and haven't tried an mcp-enabled ide yet, you are basically coding like a caveman. It is wild how fast tools like Cursor and Windsurf became the "cool kids" by letting ai actually touch your local files.

The big deal here is that these ides aren't just chatting anymore; they're acting as mcp clients. Instead of you copy-pasting code into a window, the ai uses a Filesystem mcp server to read your docs or even run a terminal command.

• github is all in: they recently put out an official mcp server. Now your ai can handle pr management or hunt through issues without you leaving the editor.
• no more context rot: because the protocol is standard, you can plug in a server for your specific database and the ai "just gets it."
• safe sandboxing: you usually pick which folders the ai can see, so it doesn't go wandering into your personal photos or ssh keys by accident.

I've seen people use the Fetch server to pull in api docs on the fly, which is a total lifesaver. Like I mentioned before, this is all about killing the friction that makes building agents such a drag. Next, we're gonna look at how databases are getting in on the action.

Database and knowledge management vendors

Ever tried explaining a complex database schema to an ai? It’s basically like teaching a cat to do taxes—frustrating and mostly pointless.

But things are getting weirdly easy now. Database and knowledge vendors are realizing that if they want their data to be "ai-ready," they can't just throw an api at you and wish you luck. They’re turning into mcp servers so your agents can actually understand what’s inside those rows and columns.

Honestly, watching companies like Notion and Confluence jump on the mcp bandwagon is a relief. It means ai can finally search internal docs without some janky scraper. But it also opens a massive door for "tool poisoning" where a bad doc tricks your ai into doing something stupid.

• vector db bridge: Vendors are using mcp to link rag (retrieval-augmented generation) directly to agents, making data "live" instead of just static chunks.
• security is messy: When you open these pipes, you need a way to stop puppet attacks. I've seen teams look at Gopher Security because they use a 4D framework—which stands for Discover, Detect, Defend, and Deter. It basically watches what the tool is actually doing in real-time, not just what the prompt says.
• instant deployment: You can basically turn a swagger or openapi schema into a secure mcp server in like, five minutes now.

It’s making the integration tax we talked about earlier feel like a bad dream. If you can plug your database straight into a model with built-in governance, why wouldn't you?

Next up, we're going to look at some of the hidden risks that come with all this connectivity.

The hidden risks of vendor adoption

Look, I love how fast mcp is making things move, but we gotta talk about the "oh crap" moments waiting in the shadows. Opening up your entire data stack to agents is basically like giving a stranger a master key to your house because they promised to do the dishes.

The reality is that every new mcp server you spin up is a potential backdoor. If a vendor server gets compromised, or an agent gets tricked by a clever prompt, your internal data is toast.

• prompt injection is real: A bad doc in a vendor's knowledge base can "poison" the tool, making your agent leak secrets or delete records.
• identity is a mess: Agents don't have fingerprints. We need way better ways to prove an agent is who it says it is before it touches a database.
• the security gap: Since mcp often runs over simple transport layers like stdio or http, we're essentially creating a massive web of p2p links that traditional firewalls aren't ready for.

Honestly, as mentioned earlier, teams are already looking at tools like gopher security to watch these executions in real-time. It’s not just about the prompt anymore; it’s about what the tool actually does once it gets the green light.

Next, we'll look at how the big cloud providers are trying to own the backend of this whole mess.

Cloud Providers and the MCP Backend

Now that everyone is building these servers, the big question is: where do they actually live? This is where the "Big Three"—AWS, Azure, and Google Cloud—are starting to flex. They don't want to just be the pipes; they want to be the host for the entire mcp ecosystem.

AWS has been particularly aggressive here. They're positioning Amazon Bedrock as a sort of "mcp hub." Instead of you running a bunch of docker containers for your mcp servers, they want you to deploy them as serverless functions. It makes sense—if your data is already in an S3 bucket, having an mcp server sitting right next to it on the same backbone reduces latency and keeps the traffic off the public internet.

• Azure's Enterprise Play: Microsoft is baking mcp support directly into Azure AI Foundry. They're focusing on the "governance" side, basically telling big banks and healthcare companies, "Hey, use mcp, but do it inside our walled garden so it's compliant."
• Google Cloud and Vertex: Google is pushing for mcp to be the standard way Gemini interacts with enterprise data. They're making it weirdly easy to turn a BigQuery table into an mcp-ready data source with just a few clicks.
• The Managed Service Trap: The risk here is that we might trade "integration tax" for "cloud tax." If you rely on a provider's proprietary mcp hosting, you might find yourself locked into their ecosystem all over again.

I've seen some startups trying to stay "cloud-agnostic" by running their mcp backends on Cloudflare Workers or Fly.io, which is a smart move if you don't want to be beholden to one giant. But for most big companies, the convenience of having AWS or Azure manage the scaling and security of their mcp servers is going to be hard to pass up.

At the end of the day, the cloud providers are racing to become the "app store" for mcp servers. Whoever wins that race basically controls the brain-to-body connection of the next generation of ai agents. Stay safe out there.