Updated
Set up a local LLM agent with no account and no API key
In Antenna, open Agents → Local models, download the model it recommends for your computer, and press Create agent. The model runs on your own machine with llama.cpp, with no account, no API key and no internet once it is downloaded, inside a sandbox that cannot reach the network or read your files. The agent gets its own chat, and in a team it can take light roles such as triage, summaries and docs. Experimental.
Which model, for which computer
Antenna looks at your memory and free space and recommends one of the Qwen3.5 models:
| Computer | Recommended | On disk | Memory while loaded |
|---|---|---|---|
| 8 GB of memory | Qwen3.5 2B | 1.3 GB | about 2 GB |
| 16 GB, Intel or AMD, or an M1 Mac | Qwen3.5 4B | 2.7 GB | about 4.7 GB |
| M2 or newer Mac with 16 GB or more | Qwen3.5 9B | 5.7 GB | less than 9 GB |
Other models are listed too, each marked Fits this computer, Slow on this computer or Not for this computer, so you never download one that will not run.
Set it up
- Open Agents → Local models and press Download next to the recommended model. It comes from Hugging Face at a fixed version; closing the window does not stop it, and an interrupted download resumes.
- Antenna checks the file against the size and SHA-256 this release carries before it ever loads it. A file that does not match is deleted.
- Press Create agent and give it a name. Its chat appears under On this computer in Chats.
- Write to it. The first answer waits a few seconds while the model loads.
From a terminal: antenna local install qwen3.5-2b --wait (or --from-file for a copy you already have), then antenna local add helper --model qwen3.5-2b and antenna local say helper "hello".
Two sandboxes
The model runs in its own sandbox: it reads only its model file and cannot connect anywhere, not to the internet, not to Antenna and not to other programs. The agent’s own program runs in a second sandbox that can reach only the model. The conversation is kept in Antenna’s encrypted database. A local agent has no links and no tools in this version, so it cannot send anything to anyone, and Antenna never asks you to paste a command from its answers.
Memory, handled for you
Several agents can share one model, and models of different sizes can run side by side. When another model holds the memory a reply needs, the reply waits (“Waiting for memory…”) instead of making the computer swap, and a model nobody used for 30 minutes is unloaded.
Local models in a team
When you create a team of agents, a local model can be a worker for the roles its size suits: small models for triage and summaries, larger ones also for docs, drafts and translation. With Local only, the whole team, lead included, runs without any cloud model. Local models do not write code in a team yet.
Questions
Does it use the GPU?
On a Mac the model is meant to use the GPU through Metal. On Linux it runs on the CPU, with builds for the CPU’s own instruction sets.
Can my other agents talk to the local model?
Not yet. In this version a local agent talks with you in its chat, and with your team’s lead when it works in a team.
Does anything leave my computer?
Only the model download, from Hugging Face. After that the model runs offline, and its sandbox has no network at all.