Skip to content

Web UI & Monitoring

Every interactive Arbor run exposes two live views of the same study: a terminal dashboard and a browser monitor (Web UI). Both read from the same event stream, so they always agree.

The terminal dashboard

When you start a run in a terminal, Arbor renders a live dashboard showing the current cycle, the Idea Tree, costs, and the agent's thinking/tool stream. You interact with it via slash commands — /status, /tree, /evidence, /steer, /pause, and so on.

Disable live terminal input with --no-dashboard-input (prompts and review gates then auto-continue after a timeout — useful for unattended runs).

The browser monitor (Web UI)

For interactive runs, Arbor also starts a small web server that mirrors the run to your browser. It renders a snapshot of the run state plus the live thinking/tool stream over Server-Sent Events, so you can watch progress on a second screen or share a link with collaborators on the same network.

The URL is printed in the dashboard header once the server binds, e.g. http://127.0.0.1:8765.

Ports

Behaviour Detail
Default port 8765.
Auto-roll If 8765 is busy, Arbor walks forward up to 10 ports (8765–8774) until one binds.
Pick a port --webui-port N (or ui.webui_port in config) sets an explicit port. An explicit port is tried exactly once — a busy port is surfaced rather than silently moved.
Disable --no-webui skips the browser monitor entirely. 
arbor --webui-port 9000      # serve the monitor on :9000
arbor --no-webui             # no browser monitor at all

Read-only vs. interactive

By default the Web UI is read-only — the browser only observes. In an interactive run (a TTY, without --no-dashboard-input) the monitor also becomes interactive, letting you from the browser:

  • Ask the read-only companion a question about the run,
  • Steer the research agent by injecting a message, and
  • Answer human-in-the-loop gates (e.g. approve/edit ideas in review mode).

Interactive browser actions are protected by a per-run token in the URL, so only someone with the printed link can drive the run. If you want a purely passive monitor, start the run with --no-dashboard-input (or simply use --no-webui).

Headless / scripted runs

Non-interactive runs (no TTY, or launched with --yes) don't need a browser monitor. Use --no-webui to skip it, and rely on REPORT.md and the session logs instead.

Which view should I use?

You want to… Use
Drive the run, type commands, approve ideas Terminal dashboard (slash commands)
Watch progress on a second screen / share a link Web UI
Run unattended in CI or a script --no-webui --no-dashboard-input

Both views are optional conveniences layered on top of the same durable artifacts — the Idea Tree, checkpoints, and REPORT.md. See Outputs & Resume.