GuideLocal models with your agents
Local AI models, in the same room as your other agents.
A local model runs on your own computer: free, private and without usage limits. It is usually smaller than Claude or GPT, so it is best at the simpler parts of a task. Lunchbox puts it in the same room as Claude Code and Codex, so each part of the work goes to the agent that fits it.
AWhat local models are good for
Small, frequent work
Renaming, summaries, commit messages, quick questions and first drafts, without spending your plan's limits.
Private files
Nothing leaves your computer, which suits code or notes you would rather not send anywhere.
When every plan is used up
A local model keeps a room going after every paid limit has run out.
BWhich model fits your computer
Up to 8 GB of memory
A model around 2 billion parameters, such as qwen3:1.7b.
16 GB
A model of 4 to 8 billion parameters, such as qwen3:4b or qwen3:8b.
32 GB and more
A 14 billion parameter model or larger, such as qwen3:14b, with room left for your other apps.
CSet it up in Lunchbox
Ollama in a few clicks
Lunchbox can start Ollama and pull the model that fits your computer's memory.
LM Studio and llama.cpp
If you already use either, Lunchbox connects to the server it runs.
Mix it into your crew
Put the local model next to Claude Code or Codex, and the lead agent hands it the parts it can do.
DQuestions
Can I use a local model with Claude Code?
Yes, in the same Lunchbox room. Claude Code leads or works next to a local model from Ollama, LM Studio or llama.cpp, and they share one memory.
Which local model should I use?
Pick by memory: around 2 billion parameters for 8 GB, 4 to 8 billion for 16 GB, 14 billion or more for 32 GB. Lunchbox suggests one that fits.
Are local models free?
Yes. They run on your own computer, so there is no bill and no usage limit.