NSO service models vs raw RESTCONF - what Cursor actually costs
Frontier LLMs can read RESTCONF. Do you still need NSO service models? A lab comparison of agent turns and context size when changing BGP AS.
19-minute walkthrough - Nokia SR OS lab with BGP route reflector, NSO service packages, and MCP.
The question sounds naive: if Claude and Cursor can read and drive RESTCONF APIs today, why bother modeling services in Cisco NSO?
I did not argue this in the abstract. I ran it on my Nokia SR OS triangle (R1/R2/R3 plus NetSim simulators). R2 is the route reflector; R1/R3 are clients. The peering config lives in NSO service packages: YANG in, XML template out.
What Cursor does without service models
I asked Cursor to walk through a concrete change: update the shared iBGP AS (65000).
Without a service model: - fetch or remember the device inventory - pull the full BGP subtree as JSON per router (Nokia YANG, NetSim, IOS-XR shapes all in chat) - PATCH each leaf individually - check sync, fix drift, retry on failure
With a service model: - one MCP call for the AS on the reflector - seven small calls for the clients - small JSON in, small JSON out - rollback = delete the service instance
Cursor’s own estimate in the session: roughly three to five times fewer agent turns on the modeled path. On three routers you barely notice. At two hundred devices it matters - not because of the euro cost per token (many teams ignore that for now), but because of context size and error rate per turn.
When a service model pays off - and when it does not
Three scenarios from the video:
A - parameter tweak: change AS for everyone. One default in the template, no per-client input field. A one-line edit in the MCP tool is enough.
B - reflector only: adjust RR AS only. Still within the same design frame.
C - eBGP instead of iBGP: that is not a parameter change. You remodel the service. Cursor can do it - but that is a larger prompt, not a quick fix.
The takeaway: NSO owns CRUD on the modeled service. Cursor’s job is modeling and the MCP tools around it - not loading the full device YANG tree into chat every time.
Lab context
The video follows Lab 1 of the NSO SROS series (IS-IS + iBGP RR provisioned entirely via MCP). This one is the architecture question after that: is the modeling effort worth it?
Stack: NSO 6.x, alu-sr NED, NetSim, custom RESTCONF MCP server (~147 tools), Cursor IDE.
More on AI consulting and network automation: guandeng.de/ai.