OMP with Local models is so good. Hats off to the dev who made this

Sandbox it, makes some weird requests, amazing harness and I lived it and I'm not sure I'm right but hits up a bunch of weird sites under the hood, I wanted to do more digging to build a report before saying anything security shits new to me

OMP is so good with all the models I’ve tried!

check out 😉

it's @_can1357! hardworking and talented

been rocking omp for 4 months now with all models I could. gotta say I am impressed with how well it works and how easy it is to customize (being a pi fork and all)

have you stopped using droid as much?

I have I recently downloaded it again to try it out

any particular reasons? i've been waiting for droid to drop the things the team has been relentlessly working towards but they've also started speeding things up for the CLI, its much faster now & less laggier / flickery overall

My favorite way to use local models

I was also impressed by how well it works with Qwen 3.8 27B at least doing initial testing it one-shots everything and with an amazing quality

@_can1357

he is @_can1357 an inspiration for me personaly.

Local plus a harness beats another Max seat.

what makes OMP good ?

which local model are you running it with. tried it on ollama a while back and my tool calls fell apart fast

Pairing OMP with local models keeps the whole workflow self-contained without having to rely on remote API calls.

I wannt try OMP, Prime Agent, Droid, Grok Build and some others

i shall run my 4.1 flash with this thanks bigman

Idk how can anyone like omp, its like a nerfed claude code with too much configurability and buggy

I also think OMP is fantastic, but the advisor system could be improved. Often it’s difficult to review the final output because if the advisor triggers new reasoning you have to scroll back up through the thread to read all the outputs.

IMO the perfect combo is OMP running in Cursor's terminal. Frontier when needed, but OMP doing daily work with local models.

How?

when you say local moedls, do you mean small models e.g. 4-bit quantised etc? been trying to tune a pi/prime-agent harness for that target size and curious. OMP seems quite heavy at first glance. thoughts?

That's really good TPS on deepseek v4

Especially with a cross family advisor model (using the constant advisor of OMP)

the ollama reply is the tell: same weights fall apart in one harness and hold in another. at that point the harness is doing more of the tool calling than the model is.