docs / start / install
Install Chief
Chief needs one thing on the machine: the app. The llama.cpp engine is bundled inside it, and the model is fetched on first run. There is no runtime to install and no account to create.
1 · The app
Download the build for your machine. Chief ships for macOS — Apple silicon and Intel — and for Windows on x64. There is no Linux build. Every build is published on the releases page, alongside the source it was built from.
Builds are not yet signed. macOS will say the app cannot be verified:
open it from the Finder context menu with Open the first time, rather than
by double-clicking. Windows SmartScreen will warn once — choose
More info then Run anyway. Signing and notarisation are not in
place yet, so this is expected rather than a sign that something is wrong.
2 · The model
First run checks for the weights and offers to fetch them: one button, a progress bar, and a download that resumes where it left off if the connection drops. The file comes from Hugging Face over HTTPS, carrying no token and no cookie, and it is the only request Chief makes to a host you did not connect yourself.
Chief chooses the model from the machine — you do not pick one. A laptop with 16 GB of memory and four or more physical cores gets the 3B; anything smaller gets the 1B, which answers less well but answers.
Offline from here. Once the weights are on disk, the Today brief, the work log and the conversation all work with the network off. Only your integrations need connectivity.
3 · Connect a tool
Settings has a card for each integration. GitHub is the one to start with, and it signs in by device code: Chief shows a short code, opens your browser with the code already filled in, and waits while you authorise it.
Outlook mail and calendar sign in through your browser as well. Linear takes a personal API key you paste; Jira registers Chief with Atlassian at sign-in and asks for no write scopes at all. Any calendar that publishes an iCalendar address — Google, iCloud, Fastmail, Outlook — works by pasting that address, with no OAuth application involved.
The Today brief
The brief opens on what is waiting on you, and it keeps two questions apart on purpose: a review somebody requested of you is a request, and your own open pull request is not. They sit under separate headings rather than in one pile.
Work log
Once GitHub is connected, Chief periodically looks for pull requests you merged, asks the local model to turn each into a one-sentence achievement, and records it. Passes are idempotent, so the same merge is never logged twice, and a pass steps aside whenever you are waiting on an answer — your question is worth more than the log being current.
Asking questions
Some questions never reach the model at all. /brief, /prep and
/log, and a written list of everyday phrasings, are answered from this
machine's own disk — which is most of the speed you will notice on a laptop. Everything
else goes to the model with the tools it needs, and streams back as it is written.
If something is wrong
“Starting…” on the setup screen
The engine is reading the weights. A couple of gigabytes takes a few seconds, and the screen re-checks itself until it resolves. This is a state, not a failure.
Answers are slow
Speed follows the hardware. The status line under an answer names the step actually running; past three seconds it adds elapsed time, and past fifteen it says outright that nothing has stalled. A smaller machine is patient rather than silent.
The engine will not start
Chief keeps the last lines the server wrote and shows them rather than guessing. The usual causes are a machine without room for the model and, on macOS 12, a missing system library that arrived in 13.3.
Where is my data?
One SQLite file in the app data directory, set to mode 0600, plus your notes as markdown in a folder you chose. Deleting the file deletes the log — there is no second copy anywhere.