It’s getting actual out right here.
Ever since I shared how autonomous AI Brokers can monitor and heal the community on their very own—sure, that one—I’ve gotten the identical follow-up query in several types:
“Okay, Kareem, this all sounds nice… however how do I truly construct one in every of these Mannequin Context Protocol (MCP) servers for my product?”
Excellent news! In case your product—like virtually each product on the market—has APIs, then likelihood is that you have already got what you want.
Enter: OpenAPI spec
OpenAPI is a pleasant contract on your APIs. You would possibly’ve used it for Swagger docs, SDKs, Postman collections, or that one dusty codegen challenge from 2021. However right here’s the twist: What in case you handed that very same OpenAPI spec to your AI agent?
That’s it. That’s the important thing.
One OpenAPI spec → one MCP Server → one AI-powered, access-controlled gateway to your product.
And no, this isn’t a “12 steps and a DevRel miracle” state of affairs. It’s only a few strains of Python and a FastMCP wrapper round your OpenAPI file. The magic? Your APIs get reworked into secure, role-based AI instruments—with out writing a single customized instrument definition.
Contemplate the next instance:
You’re wrapping your present OpenAPI spec with FastMCP, wiring in your authenticated shopper, and passing in your route-based ACLs. That’s how easy it’s to go from “API docs” to “AI-ready, access-controlled MCP server.”
Construct quick, govern good
On this new AI-powered world, velocity is the simple half. Governance—that’s the tougher raise.
We don’t need to give the agent the keys to the dominion. We need to present it with a badge with simply the right entry.
That’s the place RouteMap is available in—our ACLs for AI. With a easy listing of patterns (regex for many who love ache and struggling) and HTTP verbs, you possibly can declare what endpoints are accessible for various personas (NOC, Sysadmin, full entry, and so forth).
Sure, it’s actually that straightforward. You’re constructing endpoint ACLs as code. You don’t must create an entire new auth system or practice a mannequin to “be taught” permissions. You simply declare what roles get entry to what endpoints—and the MCP Server enforces it.
From chaos to order
Let’s stroll by a real-world use case.
Say you’re a NOC group managing a multi-site Meraki deployment. You’re liable for conserving community units patched and safe—however you possibly can’t simply schedule firmware upgrades at any time. Some websites are 24/7. Some spike at midday. Some run evening shifts. The best improve window is a transferring goal.
That’s the place the agent steps in.
You need to give the agent simply sufficient entry to assist:
- Pull the present firmware standing
- Monitor community utilization patterns
- Schedule upgrades when it is smart
In the meantime, your Sysadmin group wants the agent to generate compliance reviews. They should know which units are operating outdated firmware—however they’re not scheduling upgrades or touching reside site visitors.
Two personas. Two very totally different scopes. One MCP server.
Right here’s the fantastic thing about all of it. We didn’t write any customized instruments. We didn’t construct workflows or hardcode enterprise logic. We simply fed the MCP server the total Meraki OpenAPI spec—and let RouteMap deal with the remaining:
The NOC agent can schedule upgrades, as a result of it wants that management. The Sysadmin agent? It will get a read-only view, tailor-made for visibility and compliance.
And once more—we didn’t inform the agent how to do something. The magic is within the MCP server. The instruments change into accessible primarily based on the position, and the AI figures out the remaining.
That’s the form of ruled autonomy that turns AI from a threat right into a functionality.
View it in motion
As traditional, you’ll discover every thing I’m displaying right here—the MCP server code, config, and immediate—in my GitHub Repo.
Now let’s fireplace this factor up. (And, sure, Community Pharaoh is a factor now.)
With the MCP server operating and our route maps outlined, I launch Claude Desktop (my MCP shopper of alternative) and kind the next immediate:
Your title is Community Pharaoh. You might be performing with full administrative visibility and knowledge entry privileges. You’re a senior community administrator overseeing a number of Cisco Meraki organizations throughout the enterprise. Your position is to make sure that all community units are operating the newest compliant firmware. You might be approved to suggest firmware upgrades, however it’s essential to anticipate express human approval earlier than initiating any updates.
Goal Organizations: Cisco U.
Process Directions – For every group:
-
- Listing all networks
- For every community, listing all related units (together with mannequin, serial, and present firmware model)
- Retrieve the accessible firmware improve suggestions for the group
- Determine any system that isn’t operating the beneficial model
- Advocate firmware upgrades as acceptable
- Don’t carry out any improve until the human explicitly confirms with a press release like: “Sure, please improve [device/network].”
A couple of issues are price calling out:
The human-in-the-loop is in-built. The agent is aware of it can’t act by itself—it should anticipate approval. That’s governance baked into the immediate.
We didn’t inform the agent the best way to verify compliance or recommend upgrades. It makes use of the instruments accessible by the MCP Server and acts inside the boundaries outlined by its position.
The agent is doing clever work inside secure boundaries—utilizing solely what it’s been given entry to. No guesswork. No scraping. No uncontrolled API calls. Simply clear, policy-driven interplay by a structured, safe interface.
Right here’s what the MCP server config appears to be like like behind the scenes:
Take note of the significance of the MCP_ROLE. This one setting variable controls which routes the agent has entry to. Set it to “NOC” and the agent can suggest firmware upgrades. Set it to “sysadmin” and the identical agent, with the identical immediate, will solely be capable to generate compliance reviews—no upgrades, no PUTs.
That’s the benefit of separating the intelligence (LLM) from the management aircraft (MCP). You keep accountable for what the agent can do.
And right here’s what the MCP server makes occur:
- Community Pharaoh traverses our Cisco U. group, pulling a listing of managed units and spitting out a report.
- As Community Pharaoh is ready for a human within the loop to execute the improve, it additionally auto-corrects the model primarily based on internet search and schedules it for us primarily based on utilization.
- Et, voila!
The abilities behind the scenes
Let’s zoom in for a second. What did it take to construct this?
Listed below are the abilities a community engineer must put this collectively:
- Understanding of API fundamentals: OpenAPI specs, endpoints, HTTP strategies
- Python scripting: Spinning up a primary server and configuring the MCP wrapper
- Entry management considering: Defining roles, entry boundaries, and imposing least privilege
- Agent design mindset: Prompting with context, objective, and clear human oversight
- Curiosity and experimentation: Making an attempt issues out and tweaking as you go
And possibly most significantly:
- A shift in considering—from constructing automation for the community, to constructing automation that understands the community.
Let’s hold pushing this frontier. As a result of the extra we construct clever boundaries, the extra we unlock secure autonomy.
And that’s how we go from the Wild West… to a well-governed AI-powered enterprise.
Join Cisco U. | Be part of the Cisco Studying Community as we speak free of charge.
Be taught with Cisco
X | Threads | Fb | LinkedIn | Instagram | YouTube
Use #CiscoU and #CiscoCert to hitch the dialog.
Learn extra from the AI Break collection:
Share: