One local dashboard where AI agents hand work back to me
Polaris
Part of Life Tracker
A desktop dashboard that turns what my AI agents need from me into cards I can act on, such as questions to answer, commands to run and practice sessions, served only to my own Mac.
How it works
How what my agents need from me becomes a card I can act on. Select a layer to open it.
On screen
Three screens of Polaris. Select a part to outline it, and step through the screens.
Home
1 of 3. Home: what is running, what is waiting on me, and commands to run.
Screenshots of the public standalone copy of Polaris running on its own example data. The copy I use every day reads my agent system instead.
Problem
My agents run unattended, but some things only I can do. I answer their questions and run the commands their rules forbid them. Those requests used to sit in reports, where they were easy to miss.
Approach
One window showing the same facts as the terminal pane, as a second view of them. Every action runs the pane's own code, so a click and a keypress always do the same thing. Since 4 October the main bar has five workspaces, Home, Today, Revision, Activity and Terminal, and morning and evening routines, a calendar month view and a French coach have arrived as screens of their own.
Architecture
A small Python server loads the pane as a module and keeps no state of its own. It only listens on the Mac itself and creates a secret token each time it starts. It refuses any request that does not come from the Mac, and every change must carry that token. A build script wraps it as a Mac app with a native window, installing nothing.
What I built
Questions. One queue of everything waiting on me, with the oldest blocker first. Any agent can ask a question, and my answer can start follow-up work, so a yes launches it and a no does nothing. Finished work appears as a pin the next time I open the window, one plain summary per build rather than a pin for every stage.
Run these. When an agent reaches a step it is not allowed to take, it hands me the exact command instead. Each appears as a card with Copy, Done and Drop. A command must be a single line run from a fixed folder, or it is refused. A checker keeps marking each one as safe, unverified, done or stale, and treats anything destructive as stale. A stale command has no Copy button.
Practice sessions. Question-by-question practice, with one model call per turn, answered by Claude Code, Codex or a local model. The screen came off the main tab bar on 9 October and stays reachable. Answers are saved through the same code as the command-line version, so both leave identical records.
Terminals and the phone page. Real shells in tabs. A companion page on the phone adds Done buttons and answer boxes. Everything it writes goes through one place, nothing is matched loosely, and a tapped row says "sent", never "done".
Results
A self-test drives every action against a scratch copy, with every writer replaced by a stand-in, to prove the window and the pane behave the same. Another confirms that a request without the token, or from anywhere but the Mac, is refused.
What didn't work
The first commands handed over were multi-line blocks, and a plain shell ran the comments as commands. One had gone stale overnight and would have deleted tracked files. The Terminal tab first showed only a drawing of a terminal. Replaying saved output also replayed the terminal's own queries, which typed stray characters into the shell, so replayed output is now never answered.
Stack
- Python
- JavaScript
- WebGL2
- xterm.js
- pywebview
- tmux
How these figures are sourced
The outcome figure is agent-reported: it is taken from a test log or report written by the agents that built and ran the system. The status was read from the code repository at commit c8e90f7. Every figure and claim on this page is tied to a source file or a CV line in the site's content.