Not every piece of work needs the best model in the world. A watcher checking whether a thread has gone quiet, a triage pass sorting new items into lanes, a classifier reading a feed — that work is constant, and it adds up. It should be able to run all day without costing you anything.
So Frontier can install a model onto your own machine. In Settings → Connectors, the dashed "+" card at the end of the grid opens a Hugging Face picker. Search for a model, pick a quant — the smaller the number, the smaller the download and the less memory it needs — and Frontier fetches the runtime and the weights and writes the Connector for you. Nothing to configure, no key to paste, no account to make.
What you get is a real Connector, not a side channel. It runs threads, it backs Agents, and it answers the model calls inside Agent scripts — and because it runs here, it never counts against your plan's daily runs. You can make it the default, or leave it to a single workspace that wants it; see Choosing a connector per workspace.
It's worth being honest about where it shines. A model small enough to live on a laptop is good at reading, sorting, judging and summarising, and weaker than Claude or Codex at long multi-step code edits. Give it the steady background work and keep the hard building for the bigger models — that's the split the picker is designed around.
One thing to check after installing. Some open models can use tools and some genuinely can't — the difference is baked into the model, not something Frontier can talk it into. Where a model can't, Frontier says so plainly next to it, and the workspace's connector picker warns you before you point real work at it. A model that can only chat is still a perfectly good choice for classifying and summarising; it just can't read or edit your files, so don't give it the building.
A few practical notes. The model sits idle until something asks it a question, and unloads again after ten quiet minutes, so it isn't holding memory while you aren't using it. The first message in a conversation is slower than the rest, because the model is reading the whole context for the first time; everything after that is quick. The picker tells you how much memory a quant needs against what this machine actually has, so you can choose before you download rather than after. And your backup keeps a record of what you installed, not the multi-gigabyte weights — restore onto a new machine and the model is one click from being downloaded again.