As a more general idea, I’ve had pretty great success using LLMs to write “doctor” scripts. Similar idea to what’s going on here, instead of repeatedly giving LLMs raw data files from hardware and having it find the same issues over and over, I’ve started being careful to drive the lessons learned into automated test scripts and just instinctively running them against telemetry files every time someone sends me one, whether or not there’s something wrong (or they’re just looking for some data analysis). It’s the kind of thing I’ve always wanted to do, but under time pressure would always just knock out 20 lines of numpy code, do whatever analysis or debugging, and then move onto the next fire.
Same here. Another upside is that I have a common "library" of the solved problems (ie: tools), my sessions / project reuse. It dramatically improves the perceived results and quality (I get to what I need in a fraction of the time it took me on the beginning)
I'm confused what is divulged if you use the AI feature. Does it perform bidirectional redaction of sensitive portions of captures (IP's, MAC's, data, etc)?
I guess it is if you haven’t worked with this sort of thing before
pcap is a file format for packer captures. You can use tools to capture copies of all packets traversing a network interface, both on clients and servers running operating systems like Linux and Windows, and on network switches and routers that support this (at a minimum, with enterprise gear you can at least send a copy of all traffic from specific source and destination interfaces in both receive and transmit directions to other interfaces you connect a “sniffer” to that you then run a pcap capture on, or you can often even encapsulate the copied traffic in GRE tunnel payloads and send them directly to a remote sniffer).
This is a local web app that uses Python libraries to analyze these capture files and look for common tell tale signs of specific issues, such as a TCP connection being opened and being immediately reset, which is an excellent way to find firewalls in the path that are denying connections. You can also, providing an API key, send the capture and analysis result to an LLM to discuss things further.
A sufficiently talented network engineer can do this by hand, but can save time with this tool. End users that aren’t capable of interpreting the capture due to lack of knowledge of underlying network protocols can also get results this way without escalation.
I've worked with pcap through tcpdump and wireshark before.
>This is a local web app that uses Python libraries to analyze these capture files and look for common tell tale signs of specific issues
The problem is that your 'problem' is that it's an app to find problems. So you are just kicking the can down the road of defining your problem. You present a specific example though:
>such as a TCP connection being opened and being immediately reset, which is an excellent way to find firewalls in the path that are denying connections
So the customer has a specific problem they are trying to debug and that's when they use your tool? It's not like a proactive scan of issues?
I guess I see it, but I don't get why someone's AI Agent wouldn't just run 'apt install tcpdump' or 'winget install wireshark-cli' and run the commands themselves and interpret it.
Consider evaluating an MCP or a Skill file, the UX is like a plugin in the user's existing trusted LLM software, instead of downloading a separate tool to find the problem. I'm just having trouble imagining how someone ends up downloading this otherwise.
As a more general idea, I’ve had pretty great success using LLMs to write “doctor” scripts. Similar idea to what’s going on here, instead of repeatedly giving LLMs raw data files from hardware and having it find the same issues over and over, I’ve started being careful to drive the lessons learned into automated test scripts and just instinctively running them against telemetry files every time someone sends me one, whether or not there’s something wrong (or they’re just looking for some data analysis). It’s the kind of thing I’ve always wanted to do, but under time pressure would always just knock out 20 lines of numpy code, do whatever analysis or debugging, and then move onto the next fire.
Same here. Another upside is that I have a common "library" of the solved problems (ie: tools), my sessions / project reuse. It dramatically improves the perceived results and quality (I get to what I need in a fraction of the time it took me on the beginning)
Neat! My team and I at CableLabs did something very similar, but with more of a PNM and reactive self-fix angle. [1] [2]
We demoed it to the CEOs and CTOs of the world's largest broadband companies and unfortunately it never caught traction.
[1] https://www.cablelabs.com/wp-content/uploads/2025/03/62200_P...
[2] https://www.cablelabs.com/blog/netllm-handy-automated-networ...
AI on the gateway?
What about malware impersonating the user and installing a persistent backdoor or spy on the network?
I created something similar, incase anyone wants to check it out. No login page.
https://github.com/amanverasia/estudely-packet-explorer
I'm confused what is divulged if you use the AI feature. Does it perform bidirectional redaction of sensitive portions of captures (IP's, MAC's, data, etc)?
Another tool that solves that issue that I found: https://packetsnitch.com/
Who needs a problem detector when you've got a wife?
Can an uv environment be used in a similar way to install and use the app if docker is not chosen?
Unclear what problem this solves
I guess it is if you haven’t worked with this sort of thing before
pcap is a file format for packer captures. You can use tools to capture copies of all packets traversing a network interface, both on clients and servers running operating systems like Linux and Windows, and on network switches and routers that support this (at a minimum, with enterprise gear you can at least send a copy of all traffic from specific source and destination interfaces in both receive and transmit directions to other interfaces you connect a “sniffer” to that you then run a pcap capture on, or you can often even encapsulate the copied traffic in GRE tunnel payloads and send them directly to a remote sniffer).
This is a local web app that uses Python libraries to analyze these capture files and look for common tell tale signs of specific issues, such as a TCP connection being opened and being immediately reset, which is an excellent way to find firewalls in the path that are denying connections. You can also, providing an API key, send the capture and analysis result to an LLM to discuss things further.
A sufficiently talented network engineer can do this by hand, but can save time with this tool. End users that aren’t capable of interpreting the capture due to lack of knowledge of underlying network protocols can also get results this way without escalation.
Firewalls generally dont sent a rst when blocking syn, they just silently drop it.
>pcap is a file format for packer
I've worked with pcap through tcpdump and wireshark before.
>This is a local web app that uses Python libraries to analyze these capture files and look for common tell tale signs of specific issues
The problem is that your 'problem' is that it's an app to find problems. So you are just kicking the can down the road of defining your problem. You present a specific example though:
>such as a TCP connection being opened and being immediately reset, which is an excellent way to find firewalls in the path that are denying connections
So the customer has a specific problem they are trying to debug and that's when they use your tool? It's not like a proactive scan of issues?
I guess I see it, but I don't get why someone's AI Agent wouldn't just run 'apt install tcpdump' or 'winget install wireshark-cli' and run the commands themselves and interpret it.
Consider evaluating an MCP or a Skill file, the UX is like a plugin in the user's existing trusted LLM software, instead of downloading a separate tool to find the problem. I'm just having trouble imagining how someone ends up downloading this otherwise.