Your first conversation

The conversation is the panel on the right, and there is only one of it. No list to file things into, no new-chat button, no thread to name. You talk to the machine; it keeps track of what the two of you have been talking about.

This page assumes you have been through First run.

First, connect something to think with

Setup never asked about a model, so on a fresh install there is not one. Say anything before you connect one and the failure is at least a clear one:

No model configured — choose one in Settings → AI Models (models.yml)

It makes four attempts and gives up, about half a second end to end. Nothing hangs, and nothing keeps retrying in the background. Connect a local model is the next half-page of your life; come back here afterwards.

Ask it about itself

The best first question is one only this machine can answer. Type it into the box at the bottom of the conversation panel and press Enter.

What are you, and what machine are you running on?

What comes back on a workstation with a model running locally, abridged:

I am Atlas, a macOS workstation running on Apple Silicon hardware. Here's what
I'm running on:

Machine Specifications
- CPU: Apple Silicon (M-series chip, 20 cores)
  - Load: 18.08% (1 minute) / 15.15% (5 minutes) / 13.11% (15 minutes)
  - Model unknown (tool reports "Unknown", but hardware is Apple Silicon)
- Memory: 128 GB RAM
...
Uncertainties
- CPU model name is missing from tool output (may require
  `system_profiler SPCPUDataType`).
- Disk usage details are incomplete (run `df -h` for full breakdown).

Would you like me to retrieve additional details about the hardware or services
running on this machine?

Three things in that answer are worth naming, because they are the design and not the model’s manners.

It called itself Atlas. That is the name given during First run, and it is what the machine answers to from then on.

It went and looked. Before answering it ran a tool against the machine and read the result; the load averages above were that machine’s real load averages at that second. It did not recall them from its configuration, because its configuration does not contain them.

It said what it did not know. The tool came back without a CPU model name, and instead of producing a plausible one the answer has an Uncertainties section and the command that would settle it. Grounded answers is about why that is the whole point.

The first answer can be slow

The turn above took a little over a minute: the machine plans, runs its tool, reads the result, plans again, and only then writes. How long that takes is a fact about the hardware and the connection you chose, not about Halbert. The panel shows what it is doing while it does it.

It is one conversation, not a session

Ask a follow-up with no context and no thread to pick:

What did I just ask you?

You asked: "What are you, and what machine are you running on?"

This question initiated the current thread, and I provided an overview of my
configuration and the machine's hardware/software setup in response.

Nothing was named on either side. Halbert decides which thread a turn belongs to and the client never gets to choose, which is why there is no conversation list to keep tidy — Memory and continuity covers what survives and for how long.

Each turn in the transcript carries a Forget this control, which is the unit you actually want when you want something gone.

When it wants to run something

Ask for anything that needs a shell and the command does not simply happen.

A command it rates high risk stops the turn and waits for you, naming the rating and the reason it matched:

Execute command:
  rm -f /home/you/notes.txt

Risk Level: HIGH
Reason: Recursive or forced delete

Nothing has run at that point, and the file is still there. Approve it and the command runs and the turn picks up where it stopped. Decline and the command never runs; the turn resumes anyway and answers around it. Both paths were checked against a real file on 2026-09-21 — it survived the refusal and was deleted only by the approval.

Commands below that bar arrive in the transcript as a terminal tile — the command text, a copy button, a run button, and a skip button — which is the shape described in Stage and approve a command.

The risk rating is a net, not a fence

The rating is a pattern match over the text of the command, and its patterns are not exhaustive. Measured on 2026-09-21: rm -f <file> and rm -rf <dir> are rated high and stop for approval, while a plain rm <file> is rated medium and runs without asking — as do several other commands that destroy data just as permanently. The gate is worth having and is not worth trusting with something you cannot lose. Guardrails and approvals has the full picture.

What it will not have on day one

Most of what makes Halbert interesting is a capability it has not found yet. A fresh install reports home automation, audio and web access all off, because none of the things it probes for are present, and it has no peers paired. Each of those is one page away:

• Connect Home Assistant
• Set up voice
• Add a second machine
• Read a finding — for when the machine starts bringing you things unprompted

Next

What Halbert is if you want the shape of the thing before the details, or the Reference for the generated surface of commands, endpoints, settings and capabilities.