AI Coding Is Overstated and Underrated at the Same Time
AI-assisted coding is everywhere right now. The hype is loud. The tradeoffs are quiet.
I use Cursor at OUR+‘R Technologies, Inc., Reliability Lounge, and NullorNaN Systems. That spread of work made me ask a practical question: are we actually getting the best value from the hosted API—and its limits—or is there a better mix?
So I ran a small experiment.
What I Set Up
The setup was straightforward:
- Spun up my usual dev environment: Cursor IDE and LM Studio
- Found a Qwen3 model that fit my MacBook Pro and installed it
- Spent a few hours testing Cursor against the local endpoint
No cluster. No dedicated inference box. Just a laptop and an afternoon.
What I Found
Cursor was a bit slower on the local model, but still very usable.
That result pushed two instincts in opposite directions.
As a CTO, COO, and someone who manages CapEx and worries about infrastructure overruns, my gut said: cut the API spend. If local works, why keep paying?
As someone who lives at the keyboard and knows this work does not stray far from the wall, I wondered if that was a mistake. Convenience, model quality, and latency all matter when you are in the middle of real work.
The Answer Is Simpler Than the Debate
Cursor Pro includes a chunk of API credits. Those credits are not just for the default models—they are also what unlocks custom model endpoints.
So the practical answer is not either/or.
You can run local inference for a large share of day-to-day work and still keep API access for the tasks where hosted models are worth the cost. Best of both worlds, without pretending one path solves everything.
What Comes Later
Long term, the plan for OUR+‘R is likely a small fleet of Mac Minis or Mac Studios—clustered, with a dedicated inference stack for the org.
But when I am the one doing most of the dev work, we do not need that yet. A laptop, LM Studio, and a Cursor setup that supports both local and remote models is enough for now.
The lesson is not “cancel your API subscription.” It is: test the local path, measure what you lose, and use the tooling you already pay for to blend both until the workload justifies dedicated hardware.