Hey everybody, I’m again to exploring how agentic AI may match right into a community engineer’s workflow and develop into a useful software in our software chest.
In my weblog publish, Making a NetAI Playground for Agentic AI Experimentation, I started this journey by exploring how we will make the most of Mannequin Context Protocol (MCP) servers and the idea of “instruments” to allow our AI brokers to work together with community units by sending present instructions. In case you haven’t learn that publish but, undoubtedly test it out as a result of it’s some actually fabulous prose. Oh, and there’s some actually cool NetAI stuff in there, too. 😉
Whereas it was fascinating to see how properly AI may perceive a community engineering activity offered in pure language, create a plan, after which execute that plan in the identical means I’d, there was a limitation in that first instance. The one “software” the agent had was the power to ship present instructions to the community system. I needed to explicitly present the main points concerning the community system—particulars which are available in my “supply of reality.”
To comprehend the facility of agentic AI, NetAI must have entry to the identical data as human community engineers. For immediately’s publish, I wished to discover how I may present source-of-truth information to my NetAI agent. So, let’s dig in!
NetBox presents an MCP server
NetBox has lengthy been a favourite software of mine. It’s an open-source community supply of reality, written in Python, and accessible in varied deployment choices. NetBox has been with me via a lot of my community automation exploration; it appeared becoming to see the way it may match into this new world of AI.
Initially, I anticipated to place a easy MCP server collectively to entry NetBox information. I shortly discovered that the workforce at NetBox Labs had already launched an open-source primary MCP server on GitHub. It solely supplies “learn entry” to information, however as we noticed in my first NetAI publish, I’m beginning out slowly with read-only work anyway. Having a place to begin for introducing some supply of reality into my playground was going to considerably velocity up my exploration. Completely superior.
Including NetBox to the NetAI playground
Have you ever ever been engaged on a mission and gotten distracted by one other “cool concept?” No? I suppose it’s simply me then… 🙂
Like most of my community labs and explorations, I’m utilizing Cisco Modeling Labs (CML) to run the community playground for AI. This wasn’t the primary time I wished to have NetBox as a part of a CML topology. And as I used to be prepping to play with the NetBox MCP server, I had the thought…
Hank, wouldn’t it’s nice if there have been a CML NetBox node that may very well be simply added to a topology, and that will mechanically populate NetBox with the topology data from CML?
In fact I answered myself…
Heck yeah, Hank, that’s an ideal concept!
My thoughts instantly began understanding the main points of tips on how to put it collectively. I knew it could be tremendous straightforward and quick to knock out. And I figured different folks would discover it useful as properly. So I took a “brief detour.”
I’m certain lots of you raised your eyebrows once I stated “tremendous straightforward” and “quick.” You have been proper to be skeptical, in fact. It wasn’t fairly as straightforward or easy as I anticipated. Nevertheless, I used to be capable of get it working, and it’s actually cool and useful for anybody who desires so as to add not solely a NetBox server to a CML community but additionally have it pre-populated with the units, hyperlinks, and IP particulars from the CML topology.
I nonetheless have to compile the documentation for the brand new node definition earlier than I can publish it to the CML-Group on GitHub for others to make use of. Nevertheless, take into account this weblog publish my public accountability publish, indicating that it’s forthcoming. You’ll be able to maintain me to it.
However sufficient of the aspect monitor on this weblog publish, let’s get again to the AI stuff!
Including NetBox MCP server to LM Studio
As I discussed within the final weblog publish, I’m utilizing LM Studio to run the Massive Language Mannequin (LLM) for my AI agent regionally on my laptop computer. The primary cause is to keep away from sending any community data to a cloud AI service. Though I’m utilizing a “lab community” for my exploration, there are particulars within the lab setup that I do NOT need to be public or threat ending up in future coaching information for an LLM.
If this exploration is profitable, utilizing the strategy with manufacturing information can be the following step; nevertheless, that’s undoubtedly not one thing that aligns with a accountable AI strategy.
Cloning down the netbox-mcp-server code from GitHub was straightforward sufficient. The README included an instance MCP server configuration that offered the whole lot I wanted to replace my mcp.json file in LM Studio so as to add it to my already configured pyATS MCP server.
{
"mcpServers": {
"pyats": {
"url": "http://localhost:8002/mcp"
},
"netbox": {
"command": "uv",
"args": [
"--directory",
"/Users/hapresto/code/netbox-mcp-server",
"run",
"server.py"
],
"env": {
"NETBOX_URL": "http://{{MY NETBOX IP ADDRESS}/",
"NETBOX_TOKEN": "{{MY NETBOX API TOKEN}}"
}
}
}
}
As quickly as I saved the file, LM Studio found the instruments accessible.
There are three instruments offered by the NetBox MCP server.
- netbox_get_objects: Generic software that bulk retrieves objects from NetBox. It helps “filters” to restrict the returned objects.
- netbox_get_object_by_id: Device to retrieve a single object of any kind from NetBox given an ID.
- netbox_get_changelogs: Device to lookup audit and alter occasions
I used to be, and proceed to be, within the strategy utilized by the NetBox Labs people on this MCP server. Somewhat than offering instruments to “get_devices” and “get_ips,” they’ve a single software. NetBox’s APIs and object mannequin are properly thought out, and make a generic strategy like this potential. And it actually means much less code and improvement time. Nonetheless, it primarily provides API entry to the LLM and shifts the load for “thought” and “processing the information” again to the LLM. As Agentic AI and MCP are nonetheless very new requirements and approaches, there aren’t actual greatest practices and particulars on what works greatest in design patterns right here but. I’ll come again to this strategy and what I see as some potential downsides in a while within the publish.
I then loaded the newly launched open mannequin by OpenAI, gpt-oss, and despatched the primary question.
My first thought… Success. After which I scratched my head for a second. 10 units? Scroll again as much as the CML topology picture and rely what number of units are within the topology. Go forward, I’ll wait…
Yeah. I counted seven units, too. And if I verify NetBox itself, it additionally exhibits seven units.
So what occurred? LM Studio exhibits the precise response from the software name, so I went and checked. Certain sufficient, solely seven units’ value of data was returned. I then remembered that one of many notoriously meme-worthy failings of many AI instruments is the power to rely. Blueberries anybody?
So this changed into a pleasant teachable second about AI… AI is implausible, however it may be incorrect. And it is going to be dangerous with a few of the strangest issues. Keep vigilant, my pals. 😉
After resolving the problem with the ten units, I spent a substantial period of time asking extra questions and observing the AI make the most of the instruments to retrieve information from NetBox. Usually, I used to be fairly impressed, and gaining access to source-of-truth information shall be key to any Agentic NetAI work we undertake. If you do that out by yourself, undoubtedly mess around and see what you are able to do with the LLM and your NetBox information. Nevertheless, I wished to discover what was potential in bringing instruments collectively.
Combining source-of-truth instruments with community operations instruments
I wished to begin out with one thing that felt each helpful and fairly easy. So I despatched this immediate.
I would wish to confirm that router01 is bodily linked to the appropriate units per the NetBox cable connections. > Be aware: The credentials for router01 are: `netadmin / 1234QWer` Are you able to: 1. Test NetBox for what community units router01 is meant to be linked to, and on what interfaces 2. Lookup the Out of Band IP tackle and SSH port from NetBox, use these to hook up with router01. 3. Use CDP on router01 to verify what neighbors are seen 4. Evaluate the NetBox to CDP data.
I nonetheless needed to inform the LLM what the credentials are for the units. That’s as a result of whereas NetBox is a implausible supply of reality, it does NOT retailer secrets and techniques/credentials. I’m planning on exploring what software choices exist for pulling information from secret storage in a while.
If you’re questioning why I offered a listing of steps to sort out this downside relatively than let the LLM “determine it out,” the reply is that whereas GenAI LLMs can appear “sensible,” they’re NOT community engineers. Or, extra particularly, they haven’t been educated and tuned to BE community engineers. Possible, the long run will supply tuned LLMs for particular job roles relatively than the general-purpose LLMs of immediately. Till then, the very best observe for “immediate engineering” is to offer the LLM with detailed directions on what you need it to do. That dramatically will increase the possibilities of success and the velocity at which the LLM can sort out the issue.
Let’s take a look at how the LLM dealt with step one within the request, trying up the system connections.
At first look, this appears fairly good. It “knew” that it wanted to verify the Cables from NetBox. Nevertheless, there are some issues right here. The LLM crafted what seems to be a legitimate filter for the lookup: “device_a_name”: “router01.” Nevertheless, that’s truly NOT a legitimate filter. It’s a hallucination.
A complete weblog publish may very well be written on the rationale this hallucination occurred, however the TL;DR is that the NetBox MCP server does NOT present specific particulars on tips on how to craft filters. It depends on the LLM to have the ability to construct a filter primarily based on the coaching information. And whereas each LLM has benefited from the copious quantities of NetBox documentation accessible on the web, in all of my testing, I’ve but to have any LLM efficiently craft the proper filter for something however essentially the most primary searches for NetBox.
This has led me to begin constructing my very own “opinion” on how MCP servers must be constructed, and it entails requiring much less “guessing” from the LLMs to make use of them. I’ll most actually be again extra on this subject in later posts and displays. However sufficient on that for now.
The LLM doesn’t know that the filter was incorrect; it assumes that the cables returned are all linked to router01. This results in different errors within the reporting, because the “Thought” course of reveals. It sees each Cable 1 and Cable 4 as linked to Ethernet 0/0. The reality is that Cable 4 is linked to switch01 Ethernet0/0. We’ll see how this elements in later within the abstract of knowledge.
As soon as it has the cable data, the LLM proceeds and completes the remainder of the software’s use to assemble information.
Discovering the Out of Band IP and SSH port was easy. However the first try to run “present cdp neighbors” failed as a result of the LLM initially didn’t use the SSH port as a part of the software name. However this is a superb instance of how Agentic AI can perceive errors from MCP servers and “repair them.” It realized the necessity for SSH and tried once more.
I’ve seen a number of instances the place AI brokers will resolve errors with software calls via trial and error and iteration. In reality, some MCP servers appear to be designed particularly with this because the anticipated habits. Good error messages can provide the LLM the context required to repair the issue. Much like how we as people may react and modify once we get an error from a command or API name we ship. This is a superb energy of LLMs; nevertheless, I believe that MCP servers can and must be designed to restrict the quantity of trial and error required. I’ve additionally seen LLMs “quit” after too many errors.
Let’s check out the ultimate response from the AI agent after it accomplished gathering and processing the outcomes.
So how did it do?
First, the great issues. It accurately acknowledged that the hyperlink to switch01 from NetBox matched a CDP entry. Wonderful. It additionally referred to as out the lacking CDP neighbor for the “mgmt” swap. It’s lacking as a result of “mgmt” is an unmanaged swap and doesn’t run CDP.
It could have been actually “cool” if the LLM had observed that the system kind of “mgmt” was “Unmanaged Change” and commented on that being the rationale CDP data was lacking. As already talked about, the LLM is NOT tuned for community engineering use instances, so I’ll give it a cross on this.
And now the errors… The issue with the filter for the cable resulted in two errors within the findings. There aren’t two cables on Ethernet0/0, and the “Different unused cables” aren’t linked to router01.
Hank’s takeaways from the check
I used to be undoubtedly somewhat dissatisfied that my preliminary checks weren’t 100% profitable; that will have made for an ideal story on this weblog publish. But when I’m sincere, working into just a few issues was even higher for the publish.
AI might be downright wonderful and jaw-dropping with what it could do. Nevertheless it isn’t excellent. We’re within the very early days of Agentic AI and AIOps, and there’s a lot of labor left to do, from creating and providing tuned LLMs with domain-specific data to discovering the very best practices for constructing the very best functioning instruments for AI use instances.
What I did see on this experiment, and all my experiments and studying, is the true potential for NetAI to offer community engineers a robust software for designing and working their networks. I’ll be persevering with my exploration and stay up for seeing that potential come to fruition.
There’s a lot extra I discovered from this mission, however the weblog publish is getting fairly lengthy, so it’ll have to attend for one more installment. Whereas I’m engaged on that, let me know what you consider AI and the potential for making your day by day work as a community engineer higher.
How has AI helped you lately? What’s the very best hallucination you’ve run into thus far?
Let me know within the feedback!
Learn subsequent:
Join Cisco U. | Be part of the Cisco Studying Community immediately without cost.
Be taught with Cisco
X | Threads | Fb | LinkedIn | Instagram | YouTube
Use #CiscoU and #CiscoCert to hitch the dialog.
Share: