The assistant
The companion's built-in chat agent — a multi-turn loop with tools, running on your own model key, asking before it acts.
The companion has its own assistant: a chat window under Assistant where you ask something and it answers, uses tools along the way, and can act on your machine — with your say-so at each step.
It runs on your own model key. Connect a provider once and the assistant reuses it; connect Ollama and the whole loop stays on your machine. Until a provider is connected the assistant is inert — the window shows a Connect a model provider prompt instead of a chat box.
One turn of the loop
Get it going
Connect a model
Open Assistant and, if you have not connected a provider yet, follow the Connect a model provider prompt. See Connect a model provider for the three shapes of setup. Ollama keeps everything local.
Ask something
Type in the box at the bottom and press Enter. The reply streams in as the model produces it. Press Stop to cut a turn short — the assistant stops where it is, nothing half-done runs on.
Clear a tool when it asks
When the model wants to reach for a tool it has not been cleared for, the turn pauses and a small card appears in the stream: Allow once, Always allow, or Deny. Answer and the turn continues.
Pick the model it answers with
The gear in the top corner opens Assistant settings — choose which connected provider the assistant answers with, and which one backs its memory. Voice actions and agents reuse the same choice.
The tools it can reach
A handful of tools ship ready and work the moment a model is connected: the time, a scratch note it can jot down and read back, a web page, and a git repo. Nothing to set up.
The assistant never runs one of these on its own the first time. It stops and asks. Deny is handed back to the model as a plain result — it reads the refusal and carries on without that tool, never a crash.
Always allow is remembered; deny is never punished.
Choose Always allow and that tool is cleared from then on, no more prompts. Choose Deny and the assistant simply works around the gap. You can change your mind later — grants are revocable.
See What agents can do for the full tool surface and the consent model, and Tool for what a tool actually is.
Voice
The chat window is for typing. To speak to the assistant, use the device: hold its button and talk (push-to-talk). That path runs through a workflow — the Voice Assistant built-in captures your voice, sends it through the model, and speaks or acts on the answer.
Memory
Every conversation is a thread, kept in the rail on the left, most-recent first. Pick one up later and the assistant continues where it left off — it carries the earlier turns back into the model's context for you. Start a fresh thread any time with New chat.
Where the memory lives
Threads are stored on your Mac, in the companion's own space — the titles and ordering in one file, the turns themselves in a local database. Nothing about a conversation leaves your machine except what the model provider necessarily sees to answer a turn — and with Ollama, not even that.
If you connect an embeddings provider in Assistant settings, the assistant can also recall relevant bits from across older threads, not just the one you are in.