AIE AWS Tech Talk / Governing AI Agents at Scale: Authorization, Policy and Control with Amazon Bedrock AgentCore
Moving AI agents to production raises a new question for every enterprise: how do you secure and govern autonomous, non-human identities that call multiple tools on behalf of users?
Most teams can observe what their agents do but lack the authorization and governance layer to control them at scale.
In this Tech Talk, we will show how to treat AI agents as digital employees, governed through a centralized control plane on AWS. Through a live demo, we will walk through:
A unified catalog of agents, MCP servers and tools with the AWS Agent Registry.
Tool level allow/deny rules at the gateway with AgentCore Policy, combined with fine grained runtime decisions using Amazon Verified Permissions;
Testing policies by simulating different user personas;
Model governance: routing to model providers through AgentCore Gateway, with per caller rate limits;
Fleet wide monitoring with full visibility into every agent and tool decision;
IdP agnostic policies working across Cognito, Microsoft Entra ID and others.
You will leave with a clear blueprint to bridge the gap between governance teams (security, platform, compliance) and delivery teams (agent and application developers).
There is nothing wrong with those servers. They are popular for a reason. But MCP has grown far beyond the usual recommendations.
There are now MCP servers for things like managing your network, investigating system metrics, searching personal document archives, working with Docker infrastructure, and even giving an AI its own email inbox.
So I went looking for MCP servers that don't get nearly as much attention.
Here are 10 that I think are actually worth knowing about.
1. Synapse MCP
Synapse takes an interesting approach to coding with AI.
Instead of simply giving an AI access to your files, it builds a knowledge graph of your codebase.
It can understand relationships between functions, classes, modules, and other parts of your project.
That means you can ask questions like:
"What calls this function?"
or:
"What would be affected if I change this?"
This can be much more useful than repeatedly asking an AI to search through dozens of files.
It also supports semantic search and code-aware operations.
Why I like it: It attacks one of the biggest problems with AI coding assistants: understanding how a large codebase fits together.
2. Portainer MCP
If you use Docker, you have probably heard of Portainer. But you might not have thought about connecting it to an AI.
Portainer's MCP server lets an AI interact with your Portainer environment.
That can include working with containers, images, stacks, Kubernetes environments, and other infrastructure.
Instead of constantly switching between your terminal, Portainer and your AI assistant, you can let the AI investigate what's happening.
For example:
"Which containers are currently unhealthy?"
or:
"Show me the logs for this container."
This gets particularly interesting when you're debugging a deployment.
Why I like it: It connects AI to infrastructure you are already managing instead of giving it another generic collection of tools.
3. Tailscale MCP
Tailscale makes creating private networks between your devices incredibly easy. Its MCP server takes that one step further.
An AI agent can interact with information about your Tailscale network, including devices and network configuration.
That means you could ask questions such as:
"What machines are currently connected?"
or:
"Show me the devices on my tailnet."
For someone running multiple computers, servers, homelab machines or development environments, this can be surprisingly useful.
Why I like it: Networking is one of those areas where an AI having direct access to the right information can save a lot of time.
4. Toolbox MCP
This one is particularly interesting.
Toolbox MCP provides a collection of local, deterministic utilities.
Instead of asking an LLM to perform everything itself, you can give it tools designed to perform exact operations.
Depending on the setup, these can include things like:
PDF operations
Image manipulation
QR code generation
Hashing
Regex operations
JSON and CSV processing
Text differences
Calculations
This matters because language models aren't always great at tasks where one wrong character or number matters.
If you need an exact calculation, you don't necessarily want the model doing the calculation in its head.
You want a tool to do it.
Why I like it: It's a simple idea, but it makes an AI agent more reliable.
5. Pi-hole MCP
Pi-hole is commonly used to block advertisements and trackers at the DNS level. But troubleshooting DNS isn't always fun.
The Pi-hole MCP server gives an AI access to information from your Pi-hole installation.
You can use it to inspect DNS activity and work with local DNS records.
Imagine asking:
"What domains has this device been requesting recently?"
Instead of manually digging through the Pi-hole dashboard, your AI can help investigate it.
This is a very niche MCP server. That's also why I like it for this list.
Why I like it: It turns a very specific piece of home infrastructure into something you can interact with conversationally.
6. Paperless MCP
If you use Paperless-ngx to organize documents, this one is worth checking out.
Paperless can store and organize things like invoices, receipts, PDFs, contracts and other documents. An MCP server can give your AI access to that archive.
Now instead of manually searching through hundreds of documents, you can ask questions like:
"Find the electricity bills from last year."
or:
"Find the document containing this invoice number."
This is where MCP starts becoming more interesting than simply giving an AI access to another API.
The AI becomes a natural-language interface to your own information.
Why I like it: Your document archive becomes searchable through normal questions instead of folders and filters.
7. UniFi MCP
If you use Ubiquiti's UniFi ecosystem, this is another surprisingly useful integration.
UniFi can expose a huge amount of information about your network, Devices, clients, networks, VLANs, traffic, connectivity.
An MCP server can make that information available to an AI assistant.
For example:
"Which devices are currently connected?"
or:
"Is there anything unusual about the network right now?"
This becomes especially useful when troubleshooting a network with multiple access points, switches and clients.
Why I like it: Instead of manually jumping through several UniFi dashboards, you can ask the question directly.
8. Home Assistant MCP
Home Assistant is already one of the most powerful open-source home automation platforms.
Adding MCP makes it possible for an AI to interact with that environment.
Depending on the server and permissions you configure, the AI can inspect device states and call Home Assistant services.
So instead of:
Open Home Assistant → find the device → find the service → change the state
you could simply say:
"Turn off the lights downstairs."
Or:
"What devices are currently unavailable?"
It gets even more interesting when you combine AI with existing Home Assistant automations.
Why I like it: It shows that MCP isn't only about developers. It can become an interface for an entire environment.
9. Prometheus MCP
Prometheus is widely used for monitoring applications and infrastructure but PromQL isn't exactly something everyone wants to write manually.
An MCP server can allow an AI to work with Prometheus metrics and queries.
For example, instead of manually figuring out the right PromQL query, you could ask:
"Show me the services with unusually high CPU usage."
The AI can generate the appropriate query and investigate the returned metrics.
This can be especially useful when you're trying to figure out why an application suddenly became slow.
Why I like it: It lets AI investigate the system itself instead of only looking at the code that runs the system.
10. AgentMail MCP
This might be the most unusual one on the list. AgentMail provides email infrastructure designed for AI agents with its MCP integration, an AI can work with an actual email inbox.
That opens up workflows that aren't possible with a normal coding-focused MCP setup.
For example, an agent could receive an email, process its contents, take an action, and send a response.
You could build workflows around things like:
Customer requests
Automated notifications
AI agents communicating with services
Email-based workflows
Testing applications that send emails
The interesting part isn't simply sending an email. It's giving an AI agent an identity and communication channel.
Why I like it: It makes the AI feel less like a chatbot waiting for your prompt and more like an agent that can participate in a workflow.
You Don't Need All 10
One thing I noticed while researching these MCP servers is that the best MCP server for you depends heavily on what you actually do.
If you're a developer working with large codebases, Synapse might be much more useful than Pi-hole MCP. If you run Docker infrastructure, Portainer MCP makes more sense. If you have a homelab, Tailscale or UniFi could be far more interesting. And if you manage a large document archive, Paperless MCP might be the one that actually saves you time.
That's also the point of this list.
These aren't necessarily the 10 best MCP servers for everyone.
They're 10 MCP servers that are easy to overlook but become surprisingly useful when they match the problem you're trying to solve.
There are hundreds of other MCP servers out there.
You just have to look beyond the usual recommendations.
🚀 Keep Building Your AI Setup
If you want to keep optimizing your developer workflow and local AI setups, check out the other guides in this series:
